👉 We brought years of model post-training expertise.
👉 @emetbydeel brought leading payroll and HR compliance domain expertise.
🤝 Together, we post-trained @nvidia's Nemotron 3.5 Super to surpass Claude Opus 5.5 on a global payroll model in just 3 weeks. Co-founder and CEO @origoshen sat down with @emetbydeel's @_MatiasMichael at @NVIDIA GTC Berlin to discuss how.
Learn more about how AI21 partners with companies on the pathway to owned intelligence: www.ai21.com/blog/ai21-deel-case-study/
AI21 Labs
AI21 Labs is an Israeli company specializing in Natural Language Processing (NLP), which develops AI systems that can understand and generate natural language. More on Wikipedia
-
AI21 Labs
@AI21Labs
This media couldn't be displayed. See it on the original post. -
AI21 Labs
@AI21Labs
1/2 DPO in NeMo™ Megatron Bridge, @NVIDIA’s training library for Megatron-based models, now trains like SFT: one model on the GPUs, same memory, same step time.
DPO's reference model is frozen, so its logprobs over your dataset never change. Yet the usual setup keeps a second copy of the model and runs it at every step.
Our implementation scores the reference once (offline scoring) and saves chosen/rejected logprob sums per pair to disk. Training reads them back instead.
The best part: the file stores raw sums, with no beta applied. Beta and learning rate only enter at loss time, so one scoring run covers every hyperparameter sweep you do afterwards.
>>
AI21 Labs
@AI21Labs
2/2 Does GPU layout change your reference scores? We checked.
> Pipeline Parallel (PP) 1 vs 2: identical on all 8,320 pairs
> Tensor Parallel (DP) 4 vs 8: different on every pair
Splitting the model by layers keeps the math. Splitting each matmul reorders the floating-point sums. So you can change PP without re-scoring, but not TP.
That's why TP is part of the guardrail. A mismatched score file doesn't crash. It can quietly give you wrong margins. So the trainer checks the scoring metadata (dataset, tokenizer, sequence length, TP) against its own config and refuses to start if anything is off.
PR by @AsafGardin, contributed to @NVIDIA’s Megatron Bridge repo: github.com/NVIDIA-NeMo/Megatron-Bridge/pull/6092 -
AI21 Labs
@AI21Labs
CTO @DanielGissin breaking down how we think about post-training at today's Post-Training Gathering at @NVIDIAGTC Berlin.
Drawing on insights from our recent work @alice_dot_io, @Gong_io and @deel (results below), Daniel shared how post-trained open models offer companies the chance to save costs, establish control over their data, and build their AI edge. Full results: www.ai21.com/blog/ai21-post-trains-nvidia-nemotron-35-sup… -
AI21 Labs
@AI21Labs
Proud to have our work post-training @nvidia
Nemotron 3.5 Super VL with @alice_dot_io
recognized on stage in @JensenHuang’s keynote, kicking off @NVIDIAGTC Berlin 2026💚
Our team of Forward-Deployed Researchers post-trained recently released Nemotron 3.5 Super, surpassing GPT-6.1 Sol on task quality, while costing 25x less.
Jensen's message was clear: Post-trained open models enable companies to establish sovereign intelligence, without sacrificing frontier level quality. If you're ready to start this journey, book time with one of our training experts: ai21.com/contact-sales/This media couldn't be displayed. See it on the original post. -
AI21 Labs reposted
Google Cloud Partners
@gcloudpartners
@AI21Labs cut wait times for massive model training jobs by 83% by upgrading to Google Cloud’s Hypercomputer technology, GKE, and @nvidia H100 & H200 Tensor Core GPUs. ⚡
Together, Google Cloud and NVIDIA are helping innovators eliminate computer gridlock and accelerate model iteration. Read the full story: goo.gle/4ebqwSY -
AI21 Labs reposted
Alice
@alice_dot_io
Exciting moment @ @nvidia GTC: Jensen highlighted Alice and @AI21Labs during his keynote. 💚
-
AI21 Labs
@AI21Labs
It was a pleasure working with @googlecloud's GKE and Kueue teams on this, and we’re proud of the impact we got to have on a brilliant product - with some of our production requirements making it back into Kueue itself. More details on our blog: www.ai21.com/blog/how-ai21-manages-its-gpu-fleet-with-kueue/
Google Cloud
@googlecloud
AI21 achieved an 83% reduction in AI workload wait times. By orchestrating thousands of @nvidia H100 and H200 GPUs with Kueue on GKE and AI Hypercomputer, @AI21Labs cut high-priority job wait times from 72 hours to 12 hours.
Learn more → goo.gle/4jyuzfM -
AI21 Labs
@AI21Labs
A rented frontier model makes you one of many.
But an open model post-trained on your data, built for your workflows, and compounding with every interaction you feed back into it? That’s intelligence your competitors can never get.
Our team is at @NVIDIA GTC Berlin this week, chat with them to learn how specialized intelligence becomes your AI edge. -
AI21 Labs
@AI21Labs
1/5 Over the past three weeks, our Forward-Deployed Researchers post-trained @NVIDIAAI Nemotron 3.5 Super VL into specialized models together with @Gong_io, @deel and @alice_dot_io. Across 3 different use cases, we pushed the base model to match or surpass frontier quality, at a fraction of the cost. Results below 🧵
Post not available here, view on X
AI21 Labs
@AI21Labs
4/5 In @alice_dot_io's RL environments, built from real-world attack intelligence, we trained Nemotron 3.5 Super with SFT and DPO first, then online RL (GRPO), improving attack resilience by 41 points without harming overall task performance.
AI21 Labs
@AI21Labs
5/5 Across the board, results show the business impact and differentiation of sovereign specialized intelligence, and the opportunity for companies to reach AI independence. Thanks to @nvidia, @Gong_io, @deel and @alice_dot_io for the collaboration 🤝 Full results:
-
Post not available here, view on XPost not available here, view on X
AI21 Labs
@AI21Labs
4/5 In @alice_dot_io's RL environments, built from real-world attack intelligence, we trained Nemotron 3.5 Super with SFT and DPO first, then online RL (GRPO), improving attack resilience by 41 points without harming overall task performance.
AI21 Labs
@AI21Labs
5/5 Across the board, results show the business impact and differentiation of sovereign specialized intelligence, and the opportunity for companies to reach AI independence. Thanks to @nvidia, @Gong_io, @deel and @alice_dot_io for the collaboration 🤝
Full results: -
AI21 Labs reposted
🥑 Yet Another (AI) Yuval
@YuvalinTheDeep
How do developers actually use coding agents? 7 weeks & 4k+ turns of real developer traffic, classified by category.
Held up against public coding benchmarks, my colleague Eli Lepkifker (from @AI21Labs) found a real difference across:
- Task type: Confirmations, redirections, git commands, and pasted context make up about 40% of real developer turns but appear in almost no public coding benchmarks
- Task difficulty: Not every question is a complex, multi-file architectural issue. Sometimes it's just, 'Why don't I see my changes in git status?' Public benchmarks over-index on the former.
- Task context: Public benchmarks tasks are clean, 1k characters messages. Real developer messages are...not. We found they avg 80 characters, are sent mid-session and 60% contain more than one distinct task. -
AI21 Labs
@AI21Labs
We analyzed a year of agent research across 4 tasks: deep research, RAG indexing, agentic search and coding.
Found that first asking ‘Is this task verifiable?’ exposed inefficiencies in the agent architecture that were worth optimizing before reaching for the scaling lever.
That’s our verifiability litmus test:
> If you can't verify the task → aggregate
> If you can verify it → generate cheaply and spend your budget on the check
Full write up here: www.ai21.com/blog/the-verifiability-litmus-test-for-agent…
🥑 Yet Another (AI) Yuval
@YuvalinTheDeep
There's one question you should ask before you build an agent. Most people skip it (and pay for it in compute).
Over a year of agent research at @AI21Labs across deep research, RAG indexing, agentic search and coding, and it comes down to this: can a candidate answer be checked against anything?
Property of the task, not of your model. The answer is yes? Verify, then commit to one. No? Run varied attempts and combine them.
>> -
AI21 Labs reposted
🥑 Yet Another (AI) Yuval
@YuvalinTheDeep
There's one question you should ask before you build an agent. Most people skip it (and pay for it in compute).
Over a year of agent research at @AI21Labs across deep research, RAG indexing, agentic search and coding, and it comes down to this: can a candidate answer be checked against anything?
Property of the task, not of your model. The answer is yes? Verify, then commit to one. No? Run varied attempts and combine them.
>> -
AI21 Labs
@AI21Labs
We let open models access the outside world during evals on tasks from public benchmarks. Here's what happened:
Most went and found upstream commits containing fixes for the tasks they were supposed to solve.
And across every model, scores were higher when the agent found the fix. For some, that gap is huge:
• GPT 5.6 Luna: 0.40 → 0.72
• GLM-5.3: 0.60 → 0.84
• MiniMax M3: 0.31 → 0.63
Introduces big questions about what evals are measuring: ability to code, or ability to find an answer?
Evaluated on tasks from SWE-rebench, SWE-smith, Atlas QnA and more. -
AI21 Labs
@AI21Labs
Catch up with our two talks from @aiDotEngineer World's Fair, now available online👇
Stop chunking like it's 2022: www.youtube.com/watch?v=r9OwPx_HoV0
A vLLM debugging detective story: www.youtube.com/watch?v=btxG75rNJC4
🥑 Yet Another (AI) Yuval
@YuvalinTheDeep
My talk from @aiDotEngineer SF is up!
While everyone moved on to agents, somebody still has to take care of the chunking.
One window has to serve “what is X” and “why did X happen.” It can’t. Small cuts the reasoning, big buries the fact. So stop picking one. -
AI21 Labs reposted
🥑 Yet Another (AI) Yuval
@YuvalinTheDeep
My talk from @aiDotEngineer SF is up!
While everyone moved on to agents, somebody still has to take care of the chunking.
One window has to serve “what is X” and “why did X happen.” It can’t. Small cuts the reasoning, big buries the fact. So stop picking one. -
AI21 Labs reposted
Dr. Alex Wissner-Gross
@alexwg
Conversations in Action episode #14 with @Stanford emeritus professor and @AI21Labs co-founder Yoav Shoham (@YShoham) on "What Comes After Transformers" is out. www.youtube.com/watch?v=LVmhs87GZq0
-
AI21 Labs
@AI21Labs
Sometimes all your agents are cooking at once. So you build Claudino 🦖
A one-key terminal game to play while Claude Code is thinking, from AI21's Director of Data Isgav Davidowitz.
Think @googlechrome's no-wifi game, now for the @claudeai era.
Install code: github.com/IsgavD/Claudino
Thanks, Isgav, for improving run wait times everywhere. -
AI21 Labs
@AI21Labs
Independent verifiers improve agent output - but frontier verifiers are expensive. So we trained our own.
Results on @GoogleDeepMind's FACTS-Search:
> No verifier: 89.4 (previously published SOTA)
> With a frontier verifier: 93.4
> With our trained 8B verifier: 92.9 (at 3.2× lower cost than the frontier verifier)
Another example of when frontier-only isn't worth it.
Full write up: www.ai21.com/blog/you-need-a-verifier/ -
AI21 Labs
@AI21Labs
Adding a verifier improves every agentic AI system. But frontier verifiers are expensive - so we trained our own. Matching quality, at a fraction of the cost. Thread below 🧵
🥑 Yet Another (AI) Yuval
@YuvalinTheDeep
1/ Your agent already found the right answer, but your aggregation threw it away. Across every generator we tested (weak and frontier, open and closed), pass@k surpasses pass@1. Correct answers are routinely in the pool and just not selected. The binding constraint isn't answer generation, it's selection. (Work by @AI21Labs)
-
AI21 Labs reposted
🥑 Yet Another (AI) Yuval
@YuvalinTheDeep
1/ Your agent already found the right answer, but your aggregation threw it away. Across every generator we tested (weak and frontier, open and closed), pass@k surpasses pass@1. Correct answers are routinely in the pool and just not selected. The binding constraint isn't answer generation, it's selection. (Work by @AI21Labs)
-
AI21 Labs reposted
atomic.chat
@atomic_chat_hq
Atomic Chat signed the Open Weights letter!
We believe everyone should be able to run AI on their own device. When a model is open, thousands of teams fine-tune it, quantize it and build new tools on top of it. Together we move AI faster than any one company ever could
Jensen Huang
@JensenHuang
For my first post, I’m sharing a letter @NVIDIA signed on why open models matter.
AI will transform every industry, power every company, and be built by every country.
Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty.
The world needs both frontier closed models and frontier open models.
images.nvidia.com/pdf/Open-Weights-and-American-AI-Leader… -
AI21 Labs
@AI21Labs
AI21 joined @nvidia, @Microsoft , @a16z, and dozens of others in signing the Open Weights and American AI Leadership letter.
We build agentic systems and agent optimization products on top of models, open and closed alike. The competition between them is what keeps the field honest - on quality, on cost, on who gets access.
Both kinds of models will exist, and both should. Closing off the open path would be a mistake the whole ecosystem pays for.
Jensen Huang
@JensenHuang
For my first post, I’m sharing a letter @NVIDIA signed on why open models matter.
AI will transform every industry, power every company, and be built by every country.
Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty.
The world needs both frontier closed models and frontier open models.
images.nvidia.com/pdf/Open-Weights-and-American-AI-Leader… -
AI21 Labs
@AI21Labs
New episode drop: Thrilled to welcome Not Diamond co-founder & CEO @tomas_hk to YAAP, hosted by @YuvalinTheDeep, to unpack why the old model routing playbook breaks.
🎙️Listen here: www.ai21.com/yaap/the-model-routing-paradigm-that-changed… -
AI21 Labs
@AI21Labs
1/3 New SOTA on the full 731-task SWE-Bench Pro: 80.8% resolve at $5.99/task by building our coding agent pipeline like a team: junior + senior + principal models.
AI21 Labs
@AI21Labs
2/3 Most agent work is exploration & doesn't need frontier quality. So: juniors (open model, parallel rollouts) map where the fix lives → senior (cheap model) extracts exactly the code it depends on → principal (frontier) writes the patch, once.
AI21 Labs
@AI21Labs
3/3 Frontier is just 25% of the budget, so the pipeline runs ~⅓ the cost of a frontier agent end-to-end while beating it on quality. Well ahead of other hybrids too: closest comparison hit 66.1% (60-task subset vs. our full set). Full write-up: www.ai21.com/blog/better-and-cheaper-together-open-models…
-
AI21 Labs
@AI21Labs
Your agents are made of models + harnesses. So is your bill.
@Get_Writer's "Harness Effect" study held models fixed, swapped only the harness: cost per task dropped 41%, quality at parity.
The lesson: orchestration, not just model choice, sets the bill.
We saw that with early stopping for SWE agents: by training classifiers to spot when a candidate pool is good enough to stop generating rollouts, we cut compute by up to 44% while keeping quality flat. Full writeup: www.ai21.com/blog/improving-best-of-n-with-budget-aware-e… -
AI21 Labs
@AI21Labs
1/3 Best-of-N leaves $$ on the table by not accounting for variance in task difficulty. We built budget-aware execution: turn the dial on compute or speed, while keeping quality constant, to save costs or cut latency. It looks something like this:
AI21 Labs
@AI21Labs
2/3 By building a reliable early stopping mechanism, we could apply cascading (save up to 44% compute by not running unnecessary rollouts ) or parallel execution (up to 25% faster by sparing wait time on additional rollouts).
AI21 Labs
@AI21Labs
3/3 Best-of-N is too flat for real world variance in task difficulty. Budget-aware execution turns compute & speed into dials you tune per workload. Full write-up here: www.ai21.com/blog/improving-best-of-n-with-budget-aware-e…
-
AI21 Labs
@AI21Labs
1/3 We just landed #1 on DeepResearch Bench II without building a single new agent.
AI21 Labs
@AI21Labs
2/3 Six months ago, the best deep research agent scored ~45. Since, everyone's tried to beat this score with better agents. We went the other way.
AI21 Labs
@AI21Labs
3/3 We took 7 weak agents (ranks 7-13, none scoring >45) from the leaderboard & merged them into 1 report/task. Essentially boosting for deep research. The result: New #1 DRB II TotalScore of 64.38. Full write-up here: www.ai21.com/blog/merging-weak-agents-into-a-state-of-the…
-
AI21 Labs
@AI21Labs
1/5 Our latest Labs in Front piece: Agent pipeline order matters. By reversing a common agent recipe - scale first, enrich second - we reached SOTA on a Dec ‘25 to Mar ‘26 slice (123 issues): 60.9%.
Post not available here, view on X
AI21 Labs
@AI21Labs
4/5 Still came in ~$0.30 under Claude Code’s spend at a similar score. So we added a lightweight Test Agent that writes repo tests and filters failing patches, pushing our final result to 60.9% - surpassing Claude Code (60.9% vs 56.2%) at the same cost.
AI21 Labs
@AI21Labs
5/5 Takeaway: pipeline order is a hyperparameter. If you're already paying for parallel rollouts, reuse them - they're relevant context, not just candidate answers. Full write-up: [www.ai21.com/blog/first-scale-then-enrich-how-the-right-e…]
-
AI21 Labs
@AI21Labs
1/5 Caching carries a deterministic assumption baked in: same input, same output.
That breaks down with LLMs, and especially with agents running parallel rollout strategies like best-of-N sampling. 🧵Post not available here, view on X
AI21 Labs
@AI21Labs
4/5 The fix: stop encoding execution order, encode graph position instead. Now you get two independent controls: experiment name (controls the key) and component code (controls the prompt). Change either without invalidating the other.
AI21 Labs
@AI21Labs
5/5 What this unlocks:
- Change one prompt, serve cached results for all other components.
- A/B test any component against identical upstream outputs.
- Run best-of-N inference without N branches collapsing to the same answer.
Full design methodology here: -
AI21 Labs
@AI21Labs
1/5 We’re seeing 4 common agent optimization methods for hitting the right accuracy-cost or accuracy-latency tradeoff. We tried them all out on BrowseComp-Plus (results at end of the thread). 🧵
Post not available here, view on X
AI21 Labs
@AI21Labs
4/5 Best-of-N: Run one agent config N times in parallel → select the best trajectory. Leverages LLMs’ non-determinism - but hinges on a good eval mechanism (we use an LLM-as-a-Judge).
Can increase accuracy without linearly increasing latency - but parallel runs can add up $$.
AI21 Labs
@AI21Labs
5/5 Ensemble: Run multiple distinct agent configs in parallel → select the best trajectory. Leverages success of the portfolio vs single variants (see also: @/karpathy’s LLM Council). Can outperform best-of-N on accuracy but at lower cost & latency.
Here’s how each method did on BrowseComp-Plus: www.ai21.com/blog/maestro-deep-research-agents/ -
AI21 Labs reposted
This Week in Startups
@twistartups
A startup wants to pull magnesium from seawater without torching the environment. Another wants to take small language models to the podium for enterprise customers. Today @jason and @alex sat down with @magratheametals’s @biglithium and @AI21Labs’s @origoshen to better understand the technology behind each startup’s push to change the world.
Jason and Alex also discussed Cerebras’s latest IPO price range, ad-free TikTok, OpenAI’s PE push, and more!
0:00 Ori Goshen, CEO of AI21 joins the show
1:20 Plaud: If your work depends on conversations — interviews, meetings, calls — you need a Plaud NotePin. You can check it out at plaud.ai/twist and use code TWIST for 10% off!
2:39 How meta-models can route between LLMs for better token cost efficiency
6:01 Jamba as open-weight; Maestro as proprietary orchestration
10:08 LinkedIn Jobs - Hire right, the first time. Post your first job and get $100 off towards your job post at linkedin.com/twist
12:56 AI21 customer roster: FNAC, US tech giants, Israeli companies
19:01 Alex Grant, CEO of Magrathea Metals joins to discuss pulling magnesium from seawater
20:03 Live video of the Oakland pilot electrolyzer
20:10 Deel - Founders scale faster on Deel. Set up payroll for any country in minutes, hire anyone anywhere, get visas handled fast, and get back to building. Visit deel.com/twist to learn more.
22:01 Magnesium as a "gateway metal" for aluminum, defense, aerospace
23:38 Series A close, JV economics: $3,000/ton vs. $7,000/ton market
29:10 Sidebar bounty update: Glass Sidebar (Oliver Choy) demo
30:47 Netsuite - Get the free business guide Demystifying AI at netsuite.com/twist
31:47 Sidebar bounty update: Sidecast (Patrick Hughes) demo
36:09 Reducing scope to "real-time fact checker only" for final round
37:24 ro.co: Ro's insurance checker will let you know if your coverage includes GLP-1s for FREE. Go to ro.co/Twist for your free insurance check.
39:07 Cerebras IPO: $115–$125 → $150–$160 per share
44:19 Will OpenAI's compute commitments to Cerebras actually get funded?
49:40 OpenAI Deployment Company + Tomoro acquisition explained
53:51 Anthropic's parallel $1.5B PE joint venture with Blackstone & Goldman
56:40 OpenClaw's decline + competition from Cowork, Perplexity, Grok
1:04:22 TikTok's £3.99 ad-free subscription launches in the UK
1:09:09 Knicks sweep 76ers — Jason's playoff predictions
1:11:24 Off-duty: "There Is No Antimemetics Division" by qntm
1:13:25 Off-duty: Fall of Civilizations podcast by Paul Cooper
1:16:37 States' rights, federalism, housing supply & closing thoughts
🎥 Watch the full episode here 👇This media couldn't be displayed. See it on the original post. -
AI21 Labs
@AI21Labs
New #yaap episode out now 🎙️
@yuvalinthedeep sits down with @mikegchambers from @awsdevelopers to unpack harness engineering and why it's the reason most agents never make it to production.
🎧 Listen/Watch it now: www.ai21.com/yaap/everything-but-the-model-harness-engine… -
AI21 Labs
@AI21Labs
Take: stop searching for the “best model" or “best harness”. Model release → harness breaks → refind your optimal config.
Most teams handle this with manual local search, with poor visibility into accuracy/cost/latency tradeoffs.
What we need: automated, adaptive navigation of the tradeoff surface.
Read more: www.ai21.com/blog/maestro-deep-research-agents/ -
AI21 Labs reposted
AI21 Labs
@AI21Labs
We've been saying this for months. The best compliment you can give an AI system is that it behaves exactly as expected. Every time.
We built a campaign around it. It's called Boring AI: www.ai21.com/boring-agents/ -
Post not available here, view on X
AI21 Labs
@AI21Labs
We've been saying this for months. The best compliment you can give an AI system is that it behaves exactly as expected. Every time.
We built a campaign around it. It's called Boring AI: www.ai21.com/boring-agents/ -
AI21 Labs
@AI21Labs
Join @YuvalinTheDeep, Senior Developer Advocate at @AI21Labs, for a live webinar in partnership with @DataCamp: The Four Gaps Between Demo Agents and Production Systems.
If you're shipping agents to prod, this is where prototypes break and real systems begin.
🗓 May 6 · 11am ET
Register:www.datacamp.com/webinars/the-four-gaps-between-demo-agen… -
AI21 Labs reposted
IsraelTech
@IsraelTech
Most AI today is passive. You prompt it, it responds, you prompt again.
@origoshen , Co-CEO of @AI21Labs, draws a clear line between that reality and the agentic AI everyone's picturing. The direction is right. The timeline is genuinely uncertain.
That kind of calibration, from someone building at the frontier, is worth more than most of the takes circulating right now.
Full conversation with Ori on IsraelTech, hosted by @YoelTIsrael - link in the comments 👇This media couldn't be displayed. See it on the original post. -
AI21 Labs reposted
רברס עם פלטפורמה
@reversim
🚀 פרק חדש עם יובל פלג לוי מ @AI21Labs
• למה חילוץ נתונים מ-PDF בעברית שובר גם מודלים מתקדמים?
• הפתרון: החלפת מילים בעברית לאנגלית עם מידות פיזיות (Bounding Box) זהות, לשמירה על המבנה הוויזואלי מבלי לשבור טבלאות.
להאזנה 🎧 www.reversim.com/2026/04/2026-04-513-pdfrtl.html -
AI21 Labs
@AI21Labs
1/5 We hit SOTA performance on BrowseComp-Plus with 95.18% accuracy using AI21 Maestro’s agent optimization.
Here’s how we automated the search space and reached #1. 🧵Post not available here, view on X
AI21 Labs
@AI21Labs
4/5 But then we’re left with a search space problem. The possible agent configurations are infinite. And even if you manually optimize, what happens when a better model drops?
Or your use case data updates?
AI21 Labs
@AI21Labs
5/5 Maestro automatically searches across this space (model ensembles, scaling strategies, execution policies), maps the Pareto frontier, and surfaces the full accuracy–cost–latency tradeoff surface.
Read the full methodology here: www.ai21.com/blog/maestro-deep-research-agents/ -
AI21 Labs
@AI21Labs
Live from @DeepLearningAI conference: our CPO Or Dagan is taking the stage, explaining how we got SOTA on Browsecomp-Plus with the Maestro approach.
AI21 Labs
@AI21Labs
"Agent optimization should be automatic, efficient, observable and future-proof" (Or Dagan, 2 minutes ago)
AI21 Labs
@AI21Labs
The crowd goes wild! (Well, as much as you can in a dev conference with headphones 🎧)
-
AI21 Labs
@AI21Labs
Day 2 at AI Dev 26 in SF surrounded by the best builders.
📍Find us at booth 121.
@DeepLearningAI -
AI21 Labs
@AI21Labs
-
AI21 Labs
@AI21Labs
Day 1 of #aidevsf is live. 🙌
📍 Find us at booth 121 🎤 Catch our Chief Product Officer Or Dagan's session on Day 2: Efficiently Navigating the Agentic Action Space: Meta-Model Orchestration using AI21 Maestro.
Thanks @DeepLearningAI for having us.This media couldn't be displayed. See it on the original post. -
AI21 Labs
@AI21Labs
Tune in to Or Dagan on the Life Self Mastery podcast with @rohitmal, unpacking why most enterprise AI projects fail before they ever reach production.
You can also catch Or Dagan live at #aidev26 this week in SF with @DeepLearningAI.
Listen on:
- Spotify - open.spotify.com/episode/73VmSAPC33gcgwhrAYhojk
- iTunes - podcasts.apple.com/in/podcast/building-enterprise-ai-that…
- YouTube - www.youtube.com/watch?v=MIdVh-DgVVU&t=77s -
AI21 Labs
@AI21Labs
Are you coming to @DeepLearningAI AI's AI Dev 26 x SF next week?
Our Chief Product & Strategy Officer, Or Dagan, will be giving a talk diving into what our product and R&D teams are doing to automate finding the optimal balance between cost, accuracy, and latency for every agent that we build.
Check out the teaser below and head to Stage 2 at 4:15 pm on 4/29! -
AI21 Labs
@AI21Labs
We just hit #1 on the @huggingface BrowseComp-Plus leaderboard.
Best accuracy: 92.53%. Best recall: 88.79%. Lowest calibration error across all submissions.
Built with @AI21Labs Maestro.
huggingface.co/spaces/Tevatron/BrowseComp-Plus -
AI21 Labs
@AI21Labs
Attending #aidev26 by @DeepLearningAI?
Join @AI21Labs, @trychroma + @Baseten for a panel on optimizing modern AI systems. Drinks. Nikkei food. No fluff.
April 28 | 5PM | Kaiyō SF
Register → luma.com/7xdn3mke/ -
AI21 Labs
@AI21Labs
1/5 When we saw our Reducer (=LLM judge component in Maestro, our agentic framework, that selects the best output from parallel agent runs) consistently picking gold patches, we were sure Claude Opus 4.5 (knowledge cutoff Aug '25) had simply memorized the answers.
But then we sanity-checked on SWE-rebench with issues from Aug ‘25 - Feb ‘26 and the gold preference didn’t budge. 🤔Post not available here, view on X
AI21 Labs
@AI21Labs
4/5 We upgraded our original 3-line “be correct” prompt → a much more detailed prompt that enforced a hierarchy of constraints for correctness, regression safety, and minimality.
Basically, get the LLM-based reducer to stop being an aesthetic snob.🧑🎨
AI21 Labs
@AI21Labs
5/5 The takeaway: If your agent relies on an LLM judge for selection accuracy, measuring code quality isn’t enough; you need a measure of the model's inductive bias toward the "fingerprint" of a gold solution.
This was our blueprint. 👇 -
Post not available here, view on X
AI21 Labs
@AI21Labs
2/5 Turns out the model wasn't remembering the solution, but it was identifying "gold-like" aesthetics like minimality & clarity. Total form over function kind of scenario.
AI21 Labs
@AI21Labs
3/5 An example: In instance psf__requests-1724, the gold fix is 2 lines. Our agent’s functional fix was 8 lines. The LLM judge rejected the correct 8-liner as "messy" and "redundant," choosing a clean but **non-functional** fix instead.
🔗 See full patch in the blog: -
AI21 Labs
@AI21Labs
Catch Or Dagan, AI21's Chief Product & Strategy Officer, at #aidev26 on April 29: "An End to Manual Tinkering: Optimizing Accuracy, Cost & Latency in Real-World Agents".
We'll be at Booth 121. See you in SF! @DeepLearningAI @AndrewYNg -
AI21 Labs reposted
CTech
@Calcalistech
"Both the utopia and apocalypse scenarios ignore AI's limitations."
Prof. Yoav Shoham, one of the world's leading artificial intelligence researchers and founder of @AI21Labs, discusses hype cycles, unreliable systems, and the long road to trustworthy AI.
www.calcalistech.com/ctechnews/article/lgrnyc3bz -
AI21 Labs
@AI21Labs
Scaling model size is hitting diminishing returns. The real gains are in orchestration.
Our Co-Founder & Co-CEO @yshoham makes the case in a rare long-form profile by @Calcalistech today.
CTech
@Calcalistech
The man trying to ground the AI boom.
Prof. @yshoham, one of the world's leading artificial intelligence researchers and founder of @AI21Labs, discusses hype cycles, unreliable systems, and the long road to trustworthy artificial intelligence.
www.calcalistech.com/ctechnews/article/lgrnyc3bz -
AI21 Labs
@AI21Labs
AI Dev 26 brings the builders together. We'll be among them.
Meet us there 👇 @DeepLearningAI @AndrewYNg
DeepLearning.AI
@DeepLearningAI
3,000+ developers. Two days.
AI Dev 26 x San Francisco is where the people building AI come together to discuss and shape the future of software engineering.
Join Andrew Ng and leading experts across the field at Pier 48, April 28-29.
Tickets are selling out. Grab yours today: -
AI21 Labs
@AI21Labs
Routing every task to your largest model burns tokens, adds latency, and inflates costs.
@AI21Labs' Maestro Orchestration Meta Model (OMM) is the layer above your stack that dynamically selects the right model or tool for each step, optimizing cost, latency, and value automatically.
www.ai21.com -
AI21 Labs
@AI21Labs
Most agents guess the next move. They don't search.
Production systems need to branch, route, and kill bad paths before they compound. These gaps are where most teams quietly lose.
We mapped them in the article below. 👇 www.ai21.com/blog/mind-the-gap/ -
AI21 Labs
@AI21Labs
You're overpaying for inference. SWE-bench shows cheaper models solve the same easy problems as frontier ones.
The issue isn't model quality. It's that most systems don't route at all.
We broke down the 4 gaps breaking production AI systems: www.ai21.com/blog/mind-the-gap/ -
AI21 Labs
@AI21Labs
Congratulations to our Co-founder and Co-CEO @yshoham, named an @aaas Fellow for pioneering contributions to agentic AI. 28 years as a professor at @Stanford, now building the future of enterprise AI at @AI21Labs.
Read the full story: news.stanford.edu/stories/2026/03/three-stanford-faculty-… -
AI21 Labs reposted
vLLM
@vllm_project
Thanks to @AI21Labs for tracking down a silent uint32 overflow in vLLM's Mamba-1 CUDA kernel and contributing the fix.
Root cause: `uint32_t` stride × cache_index overflows silently at scale. Fix merged in #35275. The debugging story is worth a read.
🔗 www.ai21.com/blog/vllm-cuda-integer-overflow/ -
AI21 Labs
@AI21Labs
1/4 We hit a strange logprob mismatch while training Jamba 3B with GRPO.
Rollout logprobs and training-side recompute should match before any weight update. Ours didn’t.
That was the canary. 🧵Post not available here, view on X
AI21 Labs
@AI21Labs
4/4 The fix upstream? Two characters, basically:uint32_t → size_t
Weeks of debugging for one tiny type-width bug. Great reminder that in RL infra, the hardest part is often finding the right boundary around the failure.
👉Blog with the full investigation + the upstream vLLM PR: www.ai21.com/blog/vllm-cuda-integer-overflow/
AI21 Labs
@AI21Labs
Special shout out to to our Deep Learning Engineer and blog author @TamerGhattas911 👏
-
AI21 Labs reposted
🥑 Yet Another (AI) Yuval
@YuvalinTheDeep
1/ "Agent" is becoming a meaningless word 🤨
Every LLM pipeline is suddenly "agentic." Every wrapper is an "agent framework." The word has been stretched until it broke. -
AI21 Labs reposted
🥑 Yet Another (AI) Yuval
@YuvalinTheDeep
2/ I spent an hour with @BarakLenz, CTO at @AI21Labs, and one thesis kept coming back: we're not building agents. We're building an AI Operating System.
An OS manages resources, tracks what's running, and decides when to spawn or kill work. That's the bar. -
AI21 Labs
@AI21Labs
Day 2 at @nvidia GTC 26. Stop by booth 3103 to meet the team and learn more about what we’re building for enterprise AI.
@NVIDIAAI @NVIDIAAIDev
#nvidiagtc -
AI21 Labs reposted
Intel Capital
@intelcapital
At the event, we were joined by multiple portfolio companies for a quickfire round of pitches, including @AI21Labs, @anyscalecompute, @Articul8_AI, @bria_ai_, @centrical, @LiltHQ, @Minio, @scale_AI and @zenitysec. Thank you to our portfolio companies for your presentations, and to those in attendance!
-
AI21 Labs reposted
Intel Capital
@intelcapital
Intel Capital co-hosted the #enterpriseai for Impact and Scale event alongside @HSBC in London last week. The event included a panel on #operalationalizingai in enterprise led by @PippaChick and a keynote speech from @UCLEngineering professor Ivana Drobnjak. 🧵
-
AI21 Labs reposted
DeepLearning.AI
@DeepLearningAI
We're excited to announce more speakers joining us at AI Dev 26 x San Francisco to share how they're building and deploying modern AI systems.
Prepare to meet engineers, founders, and researchers from @GoogleDeepMind, @AMD, @ActianCorp, @Snowflake, @Replit, @AI21Labs, @FlwrLabs, @Apify, @Oumi_PBC, Okahu AI, @AnchorBrowser, @RefoldAI, @Cline, @Hackmamba, OnMemory AI, and more.
🎟️ Reserve your ticket before prices move to the next tier. Link in bio. -
AI21 Labs
@AI21Labs
Standard RAG falls apart on aggregative queries.
Things like "average ARR for companies with >1k employees" require reasoning across hundreds of docs, and vector retrieval just can't handle that reliably.
We built Structured-RAG to fix this:
- induces a schema from your corpus at ingestion
- maps every doc to structured SQL records
- translates queries to SQL at inference
📊 Also releasing two new aggregative QA benchmarks with the paper: arxiv.org/abs/2511.08505 -
AI21 Labs
@AI21Labs
Test-time compute scaling for agents isn’t just running multiple full trajectories and picking a winner. Duplicate a 50–150 step workflow and you multiply tool calls, latency, and cost.
Our take: allocate compute at high-uncertainty steps + use strong reducers to close the oracle gap.
📖 Read the research: www.ai21.com/blog/test-time-compute-swe-bench/
🎙 Watch the full YAAP podcast episode: www.ai21.com/yaap/the-agent-swarm-fallacy/ -
AI21 Labs reposted
Dylan Couzon
@DylanCouzon
Love seeing practical engineering posts. If you’re running @vLLM in production and hitting OOM or unstable load, this guide explains why workload profiling + tuned configs matter more than hardware alone. www.ai21.com/blog/scaling-vllm-without-oom/ @AI21Labs
-
AI21 Labs reposted
The Infinite Loop
@RunInfiniteLoop
A philosophical debate in open-source communities has turned into a central question: who owns foundation models? @thomas_macaulay analyzed this with @BlancheMinerva of @AiEleuther, @hstyagi of @SentientAGI and Amnon Morag of @AI21Labs: infiniteloop.media/p/who-owns-the-model-rethinking-openne…
-
AI21 Labs
@AI21Labs
1/4 Our Head of Human Alignment and Standards @JulieFadlon, PhD, breaks down what agent architectures have to learn from Psycholinguistics. Hint: modularity. 🧵
Post not available here, view on X
AI21 Labs
@AI21Labs
3/4 In a modular agent setup, a router first classifies the request as conversational or investigatory.
If investigatory, a retrieval module fetches relevant docs [eval: grounding], and a synthesis module integrates the evidence into a response [eval: faithfulness, coverage].
AI21 Labs
@AI21Labs
4/4 Read the full post here: www.ai21.com/blog/modular-intelligence-agent-orchestration/
-
AI21 Labs
@AI21Labs
1/4 Our Head of Human Alignment and Standards @JulieFadlon, PhD, breaks down what agent architectures have to learn from Psycholinguistics. Hint: modularity. 🧵
-
AI21 Labs
@AI21Labs
-
AI21 Labs
@AI21Labs
Dear AI agent,
Please don’t get creative with our Q4 reports. We might go to jail.
Your Finance team.
👉 Build Boring AI Agents with us: www.ai21.com/boring-agents/ -
AI21 Labs
@AI21Labs
1/5 As part of our work on improving the efficiency of our LLM online-RL training pipelines, we cut policy update step time by ~70% by introducing a model-agnostic padding minimization method. 🧵
Post not available here, view on X
AI21 Labs
@AI21Labs
4/5 So we took a different, model agnostic path: minimize padding before it reaches the model using truncation combined with padding-aware batching. 👇
AI21 Labs
@AI21Labs
5/5 Because this approach is model-agnostic, it applies to any architecture. Even on transformers (like Qwen2.5-7B by @alibaba_cloud, this method recovers ~90% of the gains of sequence packing, without relying on specific attention implementations.
Full breakdown + implementation details in our latest Labs in Front post: www.ai21.com/blog/padding-minimization-efficiency/ -
AI21 Labs reposted
vLLM
@vllm_project
Part 2 of our AI21 Labs x vLLM series 👇
After scaling insights, this one goes deep into debugging: how a 1-in-1000 gibberish failure in vLLM + Mamba was reproduced, traced, and fixed upstream.
Key lesson: even with correct kernel math, request classification timing can still corrupt state under memory pressure.
Thanks again to the @AI21Labs team for sharing the full investigation and fix details.
🔗 www.ai21.com/blog/vllm-debugging-mamba-bug/ -
AI21 Labs reposted
vLLM
@vllm_project
Great writeup from @AI21Labs on scaling vLLM for high-throughput, bursty workloads.
TL;DR: systematic config tuning + queue-based autoscaling = 2x throughput from the same GPUs. 🚀
Useful for anyone running vLLM in production with variable traffic patterns.
Thanks to the @AI21Labs team for publishing the full engineering writeup. 💙
🔗 www.ai21.com/blog/scaling-vllm-without-oom/
AI21 Labs
@AI21Labs
1/5 Go Big or Go OOM: The Art of Scaling vLLM 🎯.
We doubled throughput and cut latency in half-same GPUs, just better vLLM config then added smart autoscaling to handle traffic bursts.
Here's what we learned optimizing LLM-as-a-Judge for GRPO training. 🧵 -
AI21 Labs
@AI21Labs
1/5 Go Big or Go OOM: The Art of Scaling vLLM 🎯.
We doubled throughput and cut latency in half-same GPUs, just better vLLM config then added smart autoscaling to handle traffic bursts.
Here's what we learned optimizing LLM-as-a-Judge for GRPO training. 🧵Post not available here, view on X
AI21 Labs
@AI21Labs
3/5 The vertical fix: auto-tune vLLM config ⚙️ Our old settings were way too conservative. After tuning with Auto-Tune vLLM + GuideLLM we got: ~2× throughput, 2× lower latency, same GPU budget 🚀 @VLLM @Openshift
AI21 Labs
@AI21Labs
4/5 The horizontal fix: autoscale on queue depth, not GPU utilization 📊 100% GPU = efficiency, not overload. We use vllm:num_requests_waiting to trigger scale-up when the deployment can't keep up with incoming requests.
AI21 Labs
@AI21Labs
5/5 The takeaway: know your workload, tune your config, scale on metrics that reflect client experience.
These lessons apply beyond GRPO - any high-throughput vLLM deployment facing variable load can benefit.
📜Full blog post: www.ai21.com/blog/scaling-vllm-without-oom/ -
AI21 Labs
@AI21Labs
We asked our boring agents about Moltbook, the new social network for agents. They did not approve. 🦞 🚫
#boringagents #clawsdown -
AI21 Labs reposted
🥑 Yet Another (AI) Yuval
@YuvalinTheDeep
We tested this with "Oracle" experiments across multiple benchmarks:
- Index the corpus at several chunk sizes
- Let an oracle pick the best size per query (with ground truth)
Result: 20-40% better recall than ANY fixed chunk size. The optimal choice is query-dependent. >> -
AI21 Labs reposted
🥑 Yet Another (AI) Yuval
@YuvalinTheDeep
Results on MTEB: 7/8 configs improved. We also got significant improvements on FinanceBench, NarrativeQA, QMSUM and our Seinfeld datasets.
Tradeoff: You're storing the corpus K times (K = number of chunk sizes)
For most applications, storage cost < retrieval quality gains
>> -
AI21 Labs reposted
🥑 Yet Another (AI) Yuval
@YuvalinTheDeep
Tuning the chunking phase in RAG systems is notoriously difficult, and the feeling is that there's no "optimal" chunking method. Can it be that different queries just need different chunk sizes, even over the same corpus?
(Work by @AI21Labs)
>> -
AI21 Labs
@AI21Labs
1/5 Debugging vLLM: The silent corruption bug
1/1000 Jamba generations collapsed into confident gibberish during RL training. No crashes, no errors, just wrong outputs with high logprobs.
The culprit? A scheduler edge case that only triggers under memory pressure. 🧵
AI21 Labs
@AI21Labs
2/5 Used gpu_memory_utilization=0.2 to reproduce quickly, but this happens naturally when the scheduler runs out of token budget and GPU blocks get recycled. New request gets 1 token → misclassified as "decode" But num_computed_tokens=0 → should be "prefill".
Post not available here, view on X
AI21 Labs
@AI21Labs
4/5 The confusion: Scheduler saw "1 token allocated" and assumed decode. But 1 token can also mean "new request that hit token budget limits."
Now fixed: check whether it's the request's first-ever token, not the allocation size.
In vLLM v0.14.0 - update if running Mamba.
AI21 Labs
@AI21Labs
5/5 Wrote up the complete investigation-from confident gibberish to the scheduler fix. Includes the false starts (suspected CUDA bugs), the breakthrough (request tracking), and lessons for anyone running stateful models at scale.
Read the blog: www.ai21.com/blog/vllm-debugging-mamba-bug/ -
AI21 Labs
@AI21Labs
This is Linda.
25,000 regulation docs. Before lunch. No smile.
She’s a custom AI agent for banking compliance.
Consistent. Accurate. Reliable. Boring (the good kind).
Swipe to meet her.
Build boring AI solutions with AI21 Labs: www.ai21.com/boring-agents -
AI21 Labs reposted
Daniel Gissin
@DanielGissin
1/6
Introducing Dynamic Data Snoozing ⏰🛌: a straightforward way to speed up RLVR by snoozing examples that are too easy for the student. Dynamic Data Snoozing can reduce compute by up to 3x without compromising quality.
Work done @ai21labs with Yuval Globerson and @inbalmagar -
AI21 Labs reposted
Yuval Peleg Levy
@Lavuy
5/5 This really boosted scores– beating strong baselines such as GPT 5.2.
To learn about more about our method and how we wound up training our own model. Read our blog post:
www.ai21.com/blog/rtl-pdf-parsing/
@urieli17 @yoavgo @JonathanBerant @rtsarfaty
@CohenShuki -
AI21 Labs reposted
Yuval Peleg Levy
@Lavuy
1/5 Tricking Parsers: Improving RTL Parsing by Faking English: PDF document parsing is an important task for RAG.
However, there are gaps in the capabilities. Especially when it comes to underrepresented languages (Results in pink show our improvements to the SOTA).
@AI21Labs 🧵 -
AI21 Labs reposted
Shuki cohen
@CohenShuki
IBM finally joins the Mamba party, just like @AI21Labs did two years ago. As @nvidia said, "Small Language Models are the Future of Agentic AI"
-
AI21 Labs reposted
Yoav Shoham
@yshoham
Also renaming the company The Boring Company. Is that ok with you @elonmusk?
AI21 Labs
@AI21Labs
Boring AI Agents? They aren't players, they are workflow slayers:
- Oliver: Humor in beta. Grounded in data.
- Nancy: Dull in chats. Doesn’t invent facts.
- Karen: Cold and distant. Highly consistent.
In enterprise AI, boring is a feature not a flaw.
👉 See how we build predictable, accurate and consistent boring AI agents: www.ai21.com/boring-agents/ -
AI21 Labs
@AI21Labs
Boring AI Agents? They aren't players, they are workflow slayers:
- Oliver: Humor in beta. Grounded in data.
- Nancy: Dull in chats. Doesn’t invent facts.
- Karen: Cold and distant. Highly consistent.
In enterprise AI, boring is a feature not a flaw.
👉 See how we build predictable, accurate and consistent boring AI agents: www.ai21.com/boring-agents/ -
AI21 Labs
@AI21Labs
“Parallel agents are easy when they’re read-only. It gets complicated the moment they change files, send emails, or touch the real world.”
This week’s Yet Another AI Podcast unpacks the Agent Swarm Fallacy with Or Dagan (CPO, @AI21Labs): why “just add agents” breaks in practice, and how test-time compute + structured execution make scaling actually work.
🎧 Listen to the podcast here: www.ai21.com/yaap/the-agent-swarm-fallacy/
AI21 Labs
@AI21Labs
Learn more here: www.ai21.com/blog/test-time-compute-swe-bench/
-
AI21 Labs reposted
Shuki cohen
@CohenShuki
Looks super cool!
Inline with @AI21Labs hybrid models, Jamba, which was the first to introduce Mamba-based models. -
AI21 Labs
@AI21Labs
“If your AI is wrong 2% of the time, you’re not in the game.”
In his interview with @vladdoes, our Co-Founder & Co-CEO @origoshen breaks down a simple truth about enterprise AI: getting a demo to work isn’t the hard part, making it reliable in production is.
▶️ Watch the interview: youtu.be/BFMR7G1QU9wThis media couldn't be displayed. See it on the original post. -
AI21 Labs reposted
Ori Goshen
@origoshen
We built a cool extension to MCP that lets you parallelize state-changing agents without breaking things and we’re working on adding it to the official protocol. 🧵Read more here: www.ai21.com/blog/stateful-agent-workspaces-mcp/
-
AI21 Labs reposted
-
AI21 Labs
@AI21Labs
New benchmark signal: Jamba2 places high on @vectara's HHEM leaderboard for hallucination rate competitive with open models in its size class.
Check out the leaderboard: huggingface.co/spaces/vectara/leaderboard -
AI21 Labs
@AI21Labs
Meet our boring AI agents.
They don’t invent facts.
They’re dull in chats. Cold. Distant. Consistent.
Built for enterprise workflows where accurate, auditable, repeatable output at scale isn’t a nice-to-have, it’s the requirement.
👉 Build boring agents: -
AI21 Labs
@AI21Labs
1/5 We ran SWE-bench 200,000+ times to get statistical confidence in agentic evals.
The main lesson wasn’t about prompts or models.
It was: agentic evaluation is an infrastructure problem. 👇🧵Post not available here, view on X
AI21 Labs
@AI21Labs
4/5 The fix was a mindset shift: stop provisioning per run. Provision per SWE-bench instance, then reuse it.
Shared per instance: repo at the right commit, MCP server, dependencies. Unique per run: execution path, files changed, commands, patch.
We keep runs isolated with dedicated MCP client + workspaces (so parallel trajectories don’t collide).
AI21 Labs
@AI21Labs
5/5 What that unlocked:
- Failure rate: 30% → 0
- Repo downloads: 8,000+ → 500
- Provisioning: per run → once per campaign
- And the “don’t burn compute” part: split Generation (patch) from Evaluation (tests).
If evaluation fails, we retry tests, not re-generate tokens.
👉Read the blog for a deep dive: www.ai21.com/blog/scaling-agentic-evaluation-swe-bench/ -
AI21 Labs reposted
Yoav Shoham
@yshoham
Sometimes the important stuff is not the glamorous. Although this is actually quite sexy, beside being critical for the enterprise. Especially if you’re turned on by things like steerability and long context.
AI21 Labs
@AI21Labs
1/4 🚀Introducing Jamba2, a memory-efficient open source model family built for total enterprise reliability and steerability.
-
AI21 Labs
@AI21Labs
1/4 MCP works great… until you run multiple subagents on the same task and they need to write files (not just read).
Parallel reasoning is easy. Parallel writes are where things explode. 🧵Post not available here, view on X
AI21 Labs
@AI21Labs
3/4 So we added an MCP Workspace layer with 5 primitives: initialize / clone / compare / merge / delete
Protocol defines the contract; each domain picks the isolation mechanism.
AI21 Labs
@AI21Labs
4/4 For code, we implemented workspaces via git worktrees (ms spin-up, branch-per-attempt, native diff/merge).
Result: 1 → 16 parallel solution attempts on the same issue, no coordination required. Discard failed branches, merge the winner, delete the rest.
👉Read the full blog: www.ai21.com/blog/stateful-agent-workspaces-mcp/ -
AI21 Labs
@AI21Labs
1/4 🚀Introducing Jamba2, a memory-efficient open source model family built for total enterprise reliability and steerability.
Post not available here, view on X
AI21 Labs
@AI21Labs
4/4 Licensed under Apache 2.0 and available today on AI21 SaaS and Hugging Face: www.ai21.com/blog/introducing-jamba2/
-
AI21 Labs reposted
Yuval Haimovits
@yuvalhaimovits
Long-horizon agentic tasks need structured execution and smart test-time compute, not just bigger models. Great breakdown using SWE-bench and how Maestro optimizes where to spend compute for compounding gains. Worth the read.
AI21 Labs
@AI21Labs
1/6 Long-horizon agentic tasks are breaking our mental models. More tokens, bigger models, and best-of-N only go so far.
Orchestrated approach to test-time-compute scaling is what long-horizon tasks need.
Here’s what we learned using SWE-bench as a test case.
Read the blog for a deep dive: www.ai21.com/blog/test-time-compute-swe-bench/ -
AI21 Labs
@AI21Labs
1/6 Long-horizon agentic tasks are breaking our mental models. More tokens, bigger models, and best-of-N only go so far.
Orchestrated approach to test-time-compute scaling is what long-horizon tasks need.
Here’s what we learned using SWE-bench as a test case.
Read the blog for a deep dive: www.ai21.com/blog/test-time-compute-swe-bench/Post not available here, view on X
AI21 Labs
@AI21Labs
5/6 With structured execution, execution is controlled and stays on track. You can scale only individual steps that will benefit from it, instead of doing it for the entire agentic trajectory.
AI21 Labs
@AI21Labs
6/6 This is the core takeaway:
With structured plans and decision-theoretic optimization techniques, you can get a lot more with the same models, tools and compute. And as you scale, the gap widens. -
AI21 Labs
@AI21Labs
🎙️ How do you make a high quality deep research agent without burning thousands of tokens?
@tavilyai's Dean Sacoransky joins our host, @yuvalinthedeep on log-driven iteration, cutting wasted tokens, and improving quality section by section.
Watch it here: -
AI21 Labs
@AI21Labs
🎙️ In this episode of Yet Another AI Podcast, we sat down with @lindavivah from @anyscalecompute on scaling AI workloads with Ray from data processing to training, inference, and agents.
We look at Ray’s roots in RL, why tools like vLLM build on it, and when Ray vs. SaaS makes sense.
Watch the episode here: www.ai21.com/yaap/just-import-ray// -
AI21 Labs
@AI21Labs
Enterprises moved fast on AI in 2025. Only a few figured out how to scale it. The data explains why.
-
AI21 Labs
@AI21Labs
🎙️New episode of Yet Another AI Podcast
@AiImagen shares how to build real moats when everyone uses the same models and how a two-person “Commando Squad” vibe-codes in production to ship fast without risking the core product.
Tune in: www.ai21.com/yaap/everyones-got-the-same-model/ -
AI21 Labs
@AI21Labs
Vibe Agent in AI21 Maestro helps you create AI agents from a single plain-English description. It suggests purpose, validation checks, tools, and model/compute settings while explaining each step in real time.
🎥Watch the video and start building with AI21 Maestro: -
AI21 Labs
@AI21Labs
From pilots to production: what does it really take to build enterprise-ready AI?
Join our upcoming webinar, “From Probabilistic to Predictable,” and learn how to scale AI workflows with reliability and control.
👉Register: www.ai21.com/lp/webinar-from-probabilistic-to-predictable/ -
AI21 Labs reposted
Stanford HAI
@StanfordHAI
📣 @StanfordCRFM just released the 2025 Foundation Model Transparency Index, which evaluates transparency practices across major foundation model developers. This year's headline? Transparency regresses across industry, reversing last year's gains. crfm.stanford.edu/fmti/December-2025/index.html
-
AI21 Labs
@AI21Labs
AI21 Maestro is @AI21Labs' orchestration platform for building reliable end-to-end AI workflows.
It combines multi-step planning, automatic compute scaling, built-in validation, proprietary RAG, and execution graphs/scorecards to keep AI agents accurate and transparent in mission-critical use cases.
👇Check it out below: www.ai21.com/maestro/This media couldn't be displayed. See it on the original post. -
AI21 Labs
@AI21Labs
🚀 Big news for the industry: AI21 Maestro is now available for deployment in your @awscloud VPC , seamlessly integrating with LLMs delivered via @awscloud Bedrock.
With AI21 Maestro, organizations can orchestrate models, tools, and data to automate mission-critical tasks, without months of R&D work.
👉 Learn more: www.cbs42.com/business/press-releases/accesswire/1113250/… -
AI21 Labs
@AI21Labs
We took to the floor at @awscloud re:Invent to ask when people last cursed at AI.
Let’s just say… the answers did not disappoint. 🤗This media couldn't be displayed. See it on the original post. -
AI21 Labs
@AI21Labs
Live from @awscloud re:Invent in Las Vegas. We are on-site at Booth #1213 and ready to show how accurate, reliable knowledge agents are already powering complex enterprise workflows.
www.ai21.com/events/reinvent-2025/ -
AI21 Labs reposted
Together AI
@togethercompute
Together AI and @AI21Labs are partnering to make highly-optimized open-source models available to more developers.
Our models are now available on @AI21Labs for building AI agents. When providers host on Together AI, their models travel with us and get exposed to more developers through these native integrations.
Try Llama, DeepSeek, Qwen, and more on Maestro now. -
AI21 Labs
@AI21Labs
We’re taking a major step toward making AI more open, reliable & enterprise-ready.
@AI21Labs is partnering with @togethercompute to give teams seamless access to open-source & frontier models through the Maestro orchestration system.
👉 Read more: www.ai21.com/blog/ai21-together-ai-partnership/
📽️ Check out the clip to see how seamless it is: -
AI21 Labs
@AI21Labs
Heading to @awscloud #reinvent? Join @AI21Labs & @deepchecks for an AI agents meetup on Dec 4 in Las Vegas. Learn from real GenAI deployments and meet teams tackling similar challenges.
Spots are limited, register here: www.ai21.com/events/reinvent-2025/ -
AI21 Labs
@AI21Labs
We have been recognized as an *Emerging Visionary* in two @Gartner_inc® Emerging Market Quadrants: Generative AI Engineering & Generative AI Model Providers.
This reflects what we hear from the market: our customers want to partner with AI companies that offer robust frontier models plus a reliable engineering stack.
www.ai21.com/blog/ai21-emerging-visionary-gartner/ -
AI21 Labs reposted
Anyscale
@anyscalecompute
Calling all Tel Aviv AI Builders!
Don't miss our Ray meetup next week on efficient LLM inference at scale. Join @AnyscaleCompute and @AI21Labs in Tel Aviv for a night of talks on efficient inference, vLLM, and Ray.
RSVP now: lu.ma/6skfv9ob -
AI21 Labs
@AI21Labs
Heading to @awscloud re:Invent in Vegas? ♠️
@AI21Labs will be on site (Dec 1–5, Booth 1213) meeting teams working on knowledge agents and enterprise AI adoption.
Join us:
♠️ Dec 3: CTO & CIO Dinner with @team8group (apply)
♠️ Dec 4: AI Agents Meetup with @AI21Labs @deepchecks & @Port_io (sign up)
♠️ All week: Book a meeting with our team
www.ai21.com/events/reinvent-2025/ -
AI21 Labs
@AI21Labs
“AI21 Maestro’s accuracy fix for RAG’s blind spots”.
Awesome to see @PaulBaier and the @GAIinsights team highlight our work on structured retrieval and reliable enterprise chatbots in this week’s essential AI news roundup.
👇Learn more about structured RAG here: -
AI21 Labs
@AI21Labs
📡 Live from AI Dev 25 x NYC!
Our Co-founder & CEO @origoshen is on stage right now discussing why reliability is the bottleneck for agents and how to unlock their full potential.
Hosted by @DeepLearningAI_ and @AndrewYNg -
AI21 Labs
@AI21Labs
🚀 Introducing Structured RAG (S-RAG)
S-RAG transforms unstructured data into a structured, query-aware representation. It then uses formal queries over structured data at runtime, so AI21 Maestro can retrieve accurate values, ensure completeness, handle inconsistencies, and improve accuracy on aggregative queries by up to 60% with near-perfect recall for exhaustive questions.
👇Read the blog: www.ai21.com/blog/structured-rag-enterprise-accuracy/
👇Dive into the paper: arxiv.org/abs/2511.08505v1
#rag #llms #structuredrag #ai21labs -
AI21 Labs reposted
i24NEWS English
@i24NEWS_EN
"While ChatGPT is focus on consumers, we target the enterprise. Consolidating massive amounts of data, accurately." Tal Guttman from AI21Labs joins Lynn Plagmeijer on #innovanation
This media couldn't be displayed. See it on the original post. -
AI21 Labs
@AI21Labs
New episode of Yet Another AI Podcast 🎙
Yuval talks with Henry Yin (Co-founder & CTO of @agihouse_org) about “The House That Builds Builders - The Origin Story of @agihouse_org.”
If you are interested in AI builder communities and turning hacks into startups, this one’s for you.
👉Listen here: www.ai21.com/yaap/story-of-agi-house/ -
AI21 Labs
@AI21Labs
How much RAM do you need to run tiny models?
Jamba Reasoning 3B runs on just 2.25 GiB, the lightest among small models like Qwen (@Alibaba_Cloud), Llama (@Meta), Granite (@IBM), and Gemma (@GoogleDeepMind).
👉 Try Jamba Reasoning 3B yourself: huggingface.co/collections/ai21labs/jamba-reasoning-3b -
AI21 Labs
@AI21Labs
Our Jamba Reasoning 3B model took on Qwen 3 4B 2507 by @Alibaba_Cloud in the same 60K token task. Jamba Reasoning 3B finished nearly 3× faster.
🎥 Watch how it played out 👇
👉 Download Jamba Reasoning 3B here: huggingface.co/collections/ai21labs/jamba-reasoning-3b -
AI21 Labs
@AI21Labs
Next stop: AI Dev 25 x NYC 🚀
Meet the @AI21Labs team at the event hosted by deeplearning.ai and @AndrewYNg
📍 NYC
🗓️ November 14, 2025 -
AI21 Labs
@AI21Labs
We wrapped an incredible day yesterday at @githubuniverse ’25. Day 2 - here we go!
#githubuniverse @GitHubCommunity -
AI21 Labs
@AI21Labs
🕸️ The web is the world’s biggest database but scraping its data at scale is tricky. On this week’s #yaap episode, @YuvalInTheDeep talks with Rony Shalit (Chief Compliance Officer, #brightdata) about what really happens behind large-scale web scraping.
🎧 Listen on Spotify, Apple Podcasts or YouTube:
Spotify: open.spotify.com/episode/1CaHshHn4sTOfZ4pebnzAt
Apple: podcasts.apple.com/us/podcast/id1825549108?i=1000733860069
YouTube: youtu.be/QCi0AD5eNoA -
AI21 Labs
@AI21Labs
A week with the @nvidia DGX Spark and we’re already running private agents trained on private data with Jamba 3B Reasoning, our hybrid Transformer Mamba model.
Check out the video to see it in action with Omri Manor and Tal Guttman ⚡
@NVIDIAGTC @NVIDIAAIThis media couldn't be displayed. See it on the original post.
AI21 Labs
@AI21Labs
TTFT vs Input Length
AI21 Labs
@AI21Labs
Decode TPS vs Input Length
-
AI21 Labs reposted
NVIDIA AI Developer
@NVIDIAAIDev
Experience on RTX AI PCs for blazing-fast inferencing via LM Studio and Llama.cpp 👇
AI21 Labs
@AI21Labs
1/5 Releasing Jamba Reasoning 3B under Apache 2.0: Hybrid SSM-Transformer architecture that tops accuracy & speed across record context lengths. e.g. 3-5X faster than Llama 3.2 3B and Qwen3 4B at 32K tokens.
-
AI21 Labs
@AI21Labs
📅 Oct 29, 11 AM ET - Join @CohenShuki (@AI21Labs ) and @BarelTayouri (mend.io) to explore how to design and defend effective, reliable, and secure AI agents.
🔗Register: webinars.techstronglearning.com/designing-and-defending-a… -
AI21 Labs reposted
CTech
@Calcalistech
“AI will become another employee”: AI21 CEO @yshoham on the rise of hybrid organizations.
The @AI21Labs co-founder envisions a future where digital and human employees work side by side.
www.calcalistech.com/ctechnews/article/l6hiweluj -
AI21 Labs
@AI21Labs
We’re joining @jfrog's MLOps Days on Oct 22 in Tel Aviv!
Catch @CohenShuki, VP of Data at @AI21Labs, sharing how small, efficient models are powering real-world AI.
🗓️ Oct 22 | 16:00–19:00
📍JFrog HQ, Tel Aviv
events.jfrog.com/mlops-days-tel-aviv/ -
AI21 Labs reposted
Flox
@floxdevelopment
End @github Universe ‘25 with a bang! 🎉
Join us for an evening of AI demos, startup stories, and great conversations with builders shaping the future of software.
Hosted by Github for Startups, Flox, and @AI21Labs, this event brings together the community’s most inspiring teams for a night of creativity, code, and connection.
When: October 29th
Where: San Francisco Brewing Co.
We’ve got: Demos, drinks & light bites
Agenda:
6:00 PM — Doors open
7:00 PM — Demo presentations ft. Flox, GitHub for Startups, AI21 Labs, @boundaryML, and Ref
8:00 PM — Networking
8:30 PM — Wrap-up
Spots filled up fast last year, grab yours now! 🚀
Link to register in the comments -
AI21 Labs reposted
Rohan Paul
@rohanpaul_ai
AI21 releases Jamba 3B, the tiny model outperforming Qwen 3 4B and IBM Granite 4 Micro!
It achieves this by mixing Transformer attention layers with Mamba state-space layers. The Mamba parts handle long sequences efficiently without storing large attention caches, while the attention layers maintain complex reasoning ability.
This structure cuts memory use and allows smooth performance on laptops, GPUs, and even mobile devices.
Processes up to 256K tokens in one context window.
On the intelligence vs speed chart, it sits far ahead of models like Gemma 3 4B, Llama 3.2 3B, and Granite 4.0 Micro, showing higher intelligence and faster generation.
Also beats larger models such as Qwen 3 4B on reasoning quality while being smaller and faster. -
AI21 Labs
@AI21Labs
🚀 @VentureBeat spotlights Jamba Reasoning 3B, our tiny open-source model that outperforms others across reasoning benchmarks and runs securely on-device.
Read the feature 👇
venturebeat.com/ai/ai21s-jamba-reasoning-3b-redefines-wha… -
AI21 Labs
@AI21Labs
🧠 Jamba Reasoning 3B leads tiny reasoning models (Artificial Analysis).
🥇 #1 on #ifbench (52%) for instruction following
📈 21 on the @ArtificialAnlys Intelligence Index
👉Charts by @ArtificialAnlys: artificialanalysis.ai/models/open-source/tiny -
AI21 Labs reposted
SiliconANGLE
@SiliconANGLE
A121 Labs’ Jamba Reasoning 3B is a powerful tiny model that promises to transform AI economics siliconangle.com/2025/10/08/a121-labs-jamba-reasoning-3b-…
-
AI21 Labs reposted
Intel Capital
@intelcapital
In @IEEESpectrum, @kateparknews covers how @AI21Labs’ Jamba Reasoning 3B model is shifting from #open-source language models toward smaller, #decentralizedai built for #edgeai systems that run on-device. Read more below.
IEEE Spectrum
@IEEESpectrum
Why are small AI models making waves in the tech world? Models like Jamba Reasoning 3B from @AI21Labs are tiny, fast, and powerful on consumer devices. spectrum.ieee.org/small-language-models?share_id=8989478
-
AI21 Labs reposted
Avanika Narayan
@Avanika15
[10/8] minions daily ship 🚢
💻 moar local lms! beyond stoked about @AI21Labs’ latest local model — jamba reasoning 3b. this model is optimized for the edge and operate 3-5x faster than other small lms. try it in minions today 👇🏽 -
AI21 Labs reposted
IEEE Spectrum
@IEEESpectrum
Why are small AI models making waves in the tech world? Models like Jamba Reasoning 3B from @AI21Labs are tiny, fast, and powerful on consumer devices. spectrum.ieee.org/small-language-models?share_id=8989478
-
AI21 Labs
@AI21Labs
1/5 Releasing Jamba Reasoning 3B under Apache 2.0: Hybrid SSM-Transformer architecture that tops accuracy & speed across record context lengths. e.g. 3-5X faster than Llama 3.2 3B and Qwen3 4B at 32K tokens.
Post not available here, view on X
AI21 Labs
@AI21Labs
4/5 The same efficiency gains apply on mobile. Running at 16K context lengths on an iPhone 16 Pro, Jamba outputs nearly 16 tokens/second, outpacing token outputs from Llama 3.2 3B, Qwen 3 1.7B, and Phi-4 Mini. Jamba is the only one that can handle up to 64K.
AI21 Labs
@AI21Labs
5/5 Available today for download & local inference on @huggingface, @kaggle, @lmstudio, and llama.cpp. huggingface.co/ai21labs/AI21-Jamba-Reasoning-3B
-
AI21 Labs reposted
World Summit AI: We do AI different.
@WorldSummitAI
Thrilled to join forces with @AI21Labs as an Exhibition Partner at World Summit AI in Amsterdam! 🚀
🎟 Use code: AI21LABS20 for 20% off tickets!
Amsterdam, 8–9 October at Taets Art and Event Park
worldsummit.ai/tickets/
#worldsummitai #wsai #aibrains #ai21 #enterpriseai #genai -
AI21 Labs
@AI21Labs
🚀 See you next week at @WorldSummitAI in Amsterdam (Oct 8–9)!📍 Booth G44 – Meet the @AI21Labs team and see how we’re shaping the future of enterprise AI.
-
AI21 Labs reposted
JFrog
@jfrog
📣 Calling all #ml pros in Tel Aviv: Join us for the free MLOps Days event at JFrog HQ on 22 October.
Learn to scale #ai, advance your ML workflows, and meet the best and brightest minds w ith talks from leaders at AI21Labs, JFrog, TensorOps, and more!
👉🏽 It's free, but seats are limited. Grab yours now: events.jfrog.com/jfrog-beyond-tokens-meetup-june-2026/ -
AI21 Labs
@AI21Labs
💻 54% of tech firms are in “Reshape” mode, redesigning workflows with AI (@BCG, 2025). From automating RFP reviews to powering customer assistants, success depends on AI that is accurate, secure, and customizable.
That’s where @AI21Labs helps technology leaders reshape workflows.
👇Explore more: www.ai21.com/industries/tech/ -
AI21 Labs
@AI21Labs
After all the talks here at @agihouse_org, all the teams are making AI21 Maestro work for them 🤓😎Thanks for @GoogleCloudTech for being our awesome partner 😍
-
AI21 Labs
@AI21Labs
We’re live at the @agi_house Hackathon: Productivity Agent Build Day 🚀 Kicking things off with the keynote — big energy in the room as builders gather to create with @AI21Labs, @GoogleCloud + @BrightData. Let’s go!
-
AI21 Labs
@AI21Labs
Everyone in healthcare is “exploring AI.” 47% of organizations (NVIDIA, 2025).
But exploring isn’t the same as executing.
What’s stopping pilots from going to production?
👉 www.ai21.com/industries/healthcare/ -
AI21 Labs
@AI21Labs
📢 NEW: You can now run Mamba-based architectures (shout-out to @tri_dao & @_albertgu for introducing Mamba), as well as hybrid Transformer–Mamba models (like our own #jamba), with the enhanced vLLM v1 engine.
Kudos to our very own @AsafGardin and Amir Koblyansky for their contribution to the #vllm project. More details on @github: github.com/vllm-project/vllm/releases
#vllm #localinference #mamba -
AI21 Labs
@AI21Labs
📢Bay Area devs! 🚀 Join us Sept 20, 10 AM PT at @agihouse_org (Hillsborough, CA) for the Productivity Agent Build Day Hackathon.
Partners: @agihouse_org @Google @googlecloud @EudiaAI & Bright Data.
👉app.agihouse.org/events/productivity-agent-20250920 -
AI21 Labs reposted
DeepLearning.AI
@DeepLearningAI
The stage for AI Dev 25 NYC is taking shape!
Announcing the first speakers you’ll hear from on November 14, and some of the partners powering the second edition of the AI Developer Conference, hosted by Andrew Ng and deeplearning.ai.
Tickets are selling fast, secure yours now ➡️ ai-dev.deeplearning.ai
A big thank you to @genspark_ai, @awscloud, @DEShawGroup, @Arm, @neo4j, @Redisinc, @AI21Labs, @coderabbitai, @LandingAI, @Snowflake, @CodioHQ, @MistralAI, @domynai, @datadoghq, @baseten, @CopilotKit, @Prolific, @AnthropicAI, Okahu AI, @tavilyai, and @giskard_ai. -
AI21 Labs
@AI21Labs
Financial institutions face rising costs, tighter margins & nonstop regulation. 👇
@AI21Labs helps automate the toughest workflows:
📊 Transform financial records into structured data
🛡️ Monitor compliance in real time
🤝 Accelerate M&A due diligence
📈 Connect macrotrend signals to strategy
www.ai21.com/industries/finance/ -
AI21 Labs
@AI21Labs
🚀 The SF/LA Tech Week 2025 calendar is live!
We’re excited to host The Next Layer of Reasoning – Finding the Missing Piece with @AITinkerers and @techweek_
📍 SoMa, San Francisco
🗓 Thursday, Oct 9 · 6:00pm
👉Sign up here: partiful.com/e/Q8ZjdwkPxvq5LBRyfyIx?accept-cohost=219afe6…
#sftechweek -
AI21 Labs
@AI21Labs
LLMs are brilliant ~95% of the time.
But in enterprise, it’s the other 5% that matters. Even small inaccuracies can carry big risks.
🎙️ Our Co-Founder & Co-CEO, Professor @yshoham explains why accuracy is the blocker and why the path forward is system-level AI.
👉 Full lecture here: youtu.be/mbGutn6-9aI -
AI21 Labs
@AI21Labs
“AI and the Future of Work: From Hype to Real Impact” - we tackled this big question with some of the sharpest voices in #enterpriseai:
✨ @yshoham – Co-Founder & Co-CEO, @AI21Labs
✨ @erikbryn – Director, @DigEconLab
✨ @nxthompson – CEO, @TheAtlantic
Key insights:
🔹 The reality of AI today vs. the hype
🔹 Why enterprises struggle to scale pilots
🔹 The future of human + AI collaboration
🎥 Watch the full replay 👇 www.ai21.com/events/ai-and-the-future-of-work/ -
AI21 Labs
@AI21Labs
🤔 That AI answer sounds brilliant… but can you trust it?
This week on Yet Another AI Podcast, @yuvalinthedeep talks with @NoamGat about judge models: reward models, critic models & how LLMs can rate, rank & critique each other.
They cover:
⚖️ Scoring vs feedback
🤖 Using judge models in inference
🚨 Why benchmarks miss the mark
🎧 Listen now on Spotify and Apple Podcasts👇
Spotify: shorturl.at/rcuRA
Apple Podcasts: shorturl.at/BAAVF
#ai21labs #yaap #ai #llms #enterpriseai -
AI21 Labs
@AI21Labs
👩💻 Calling all dev advocates! We’re #hiring a Developer Relations Engineer based in the #bayarea.
Join #ai21labs and help developers around the world build with generative AI.
Apply now 👇
shorturl.at/mygT4
#ai21labs #devrel #developercommunity -
AI21 Labs
@AI21Labs
Only 5% of generative AI pilots scale successfully.
Recent @MIT research by Aditya Challapally via @Fortune @sherylestrada, finds that many tools like #chatgpt work for individuals, but stall in the enterprise. The 5% that succeed use agentic AI systems and partner with specialized vendors.
We believe the key is implementing AI systems that ‘learn, remember, and act independently within set boundaries.’
What do you think? Join the debate 👇
fortune.com/2025/08/18/mit-report-95-percent-generative-a…
#enterpriseai #aistrategy #genai #agenticai #futureofwork -
AI21 Labs
@AI21Labs
You’ve seen hundreds of agent maps. Most are too broad.
Enterprises don’t need RAG chatbots. They need knowledge agents that handle multi-step, high-stakes tasks.
This map highlights the current players in the space. Check it out 👇
@awscloud @databricks @langchain @Microsoft @MongoDB -
AI21 Labs
@AI21Labs
Attention was never enough.
The hybrid LLM era is here—and it’s moving fast.
From Mamba to Jamba to Bamba, we mapped every major model that’s challenged the Transformer default in the past 18 months.
🧵 A timeline of what’s changed and why it matters ↓
🔗Post not available here, view on X
AI21 Labs
@AI21Labs
We broke it all down:
📚 Key papers & model architectures
🧠 Design tradeoffs: MoE, GQA, layer ordering
📊 Benchmarks across RULER, MMLU, ARC, HumanEval
🔓 Open weights + distillation strategies
Read the full story here: www.ai21.com/blog/rise-of-hybrid-llms/
AI21 Labs
@AI21Labs
FYI @gradientpub, @bensbitesdaily, @tldrnewsletter, @tldrdan, @AiBreakfast
-
AI21 Labs
@AI21Labs
Karpathy’s leash isn’t a shackle, it’s how enterprises learn to trust AI. Do you agree @karpathy ? We explore Karpathy’s idea of “putting AI on a leash” here: www.ai21.com/blog/karpathys-leash/ 🧵 1/6
Post not available here, view on X
AI21 Labs
@AI21Labs
It’s time to rethink how we architect AI systems. They should guarantee fulfillment of explicit user requirements, know how to break down work into the right steps, use the optimal tools to execute each step, and validate the output of each step. 5/6
AI21 Labs
@AI21Labs
These steps should be part of a visible, auditable plan, an “auditable artifact”, as Karpathy says, that helps users and builders understand exactly how the agent tackled the work. 6/6
-
AI21 Labs
@AI21Labs
Complex AI, Simple Assembly.
We believe that even the most complex AI systems should be easy to assemble & incredibly useful. That’s why Maestro AI agents are simple to build & bring real value to businesses.
@AI21Labs' Maestro is now available to all enterprise customers!
#enterpriseai #ai21maestro #complexaisimpleassembly @RaiseSummit -
AI21 Labs
@AI21Labs
Now live. A new update to our Jamba open model family 🎉
Same hybrid SSM-Transformer architecture, 256K context window, efficiency gains & open weights.
Now with improved grounding & instruction following.
Try it on AI21 Studio or download from @huggingface 🤗
More on what we improved & why in the release notes: docs.ai21.com/changelog -
AI21 Labs
@AI21Labs
June was a busy developer-first month for us!
- We hosted three AI Tinkerers meetups in SF, Paris, and NYC
- Presented insights about agents at the AI Engineer World's Fair conference in SF
- Launched a 7-day virtual hackathon with AWS Heroes using our AI21 Maestro product
- Aired new episodes of AI21 YAAP, developer to developer podcast (Spotify, YouTube)
Working directly with capable experts is incredible! We have more news coming soon... -
AI21 Labs reposted
Robert Scoble
@Scobleizer
It's not every day I get to interview a former principal scientist who worked at Google, and is a Professor Emeritus at Stanford University, about the state of AI.
But here we go. Introducing an hour with Yoav Shoham, @yshoham, AI pioneer and cofounder of @AI21Labs .
This will make you smarter, not that all my videos aren't that way. :-)
++++++++++++++++++
Here's what we discussed (this part was written by Chat GPT after I gave it the transcript of the video):
🚀 The State of AI Today
•The pace of AI development is unprecedented, likened to a “universal firehose” of innovation.
•Everyone—from your plumber to enterprise CTOs—is using AI. But not all use cases are equal or enterprise-ready.
🏢 Enterprise vs Consumer AI
•Enterprise adoption is still slow compared to consumer. Shoham cites AWS data showing only 6% of AI pilots go into production.
•Enterprises demand reliability, cost control, and explainability, which raw LLMs like ChatGPT don’t fully offer out of the box.
🧱 Beyond the LLM Hype
•Shoham explains that pure LLMs aren’t enough. Enterprises need “compound AI systems” or “AI agents” that:
•Use tools like calculators for arithmetic instead of relying on the model
•Integrate with company databases via RAG (retrieval-augmented generation)
•Plan, reason, and execute tasks through orchestrated workflows
•AI21 Labs built Maestro, their orchestration system, to do exactly this.
🔐 Enterprise Concerns
•Enterprises worry about IP leakage, data privacy, and hallucinations.
•AI21 addresses this by running models on-prem or in VPCs, ensuring data doesn’t leave customer control.
📉 Why Models Still Fail
•LLMs generate “authoritative bullshit” — convincing but wrong answers.
•Shoham says “prompt-and-pray” doesn’t work for serious business tasks.
•Real-world enterprise deployments need robust evaluation frameworks, not just leaderboards.
📊 Case Study: French Retailer Auchan
•Auchan deployed AI21’s system to automatically generate product descriptions—a clear ROI, but required careful iteration to build trust.
🧰 What’s Next in AI21’s R&D
•Working on planning systems, action models, and ways to estimate cost/accuracy trade-offs before running tasks.
•Focused on enterprise AI orchestration, not flashy multimodal generation.
⚠️ Agent Washing Warning
•Shoham warns against the buzzword “agent” being overused. His advice:
“Translate ‘AI agent’ to ‘software system that does X.’ If it still makes sense, keep going.”
🤖 The Human-AI Hybrid Future
•Shoham sees a world of hybrid teams: humans and AI agents working together.
•This transformation will affect everything from org charts to HR policies.
•The AI-powered worker is scalable, reliable, and multilingual — changing customer service, operations, and more.
🗣️ Closing Thoughts
•Enterprise leaders need to move beyond the fear and hype to start small, test carefully, and scale based on value.
•“AI won’t replace humans,” Shoham says, “but humans using AI will replace those who don’t.”This media couldn't be displayed. See it on the original post. -
AI21 Labs
@AI21Labs
Yesterday @origoshen our CEO Introduced Maestro, AI21’s agent framework focused on consistency and accuracy, highlighting that while 90% of enterprises aim to adopt AI across hundreds of use cases, only 1-6% reach production—mainly due to immature technology and fear of costly errors. AI21’s infrastructure enables performance optimization while ensuring outputs meet predefined conditions.
-
AI21 Labs
@AI21Labs
🚀 What a morning! ☀️☕️
Today, we hosted a packed house of hundreds at the AI Breakfast Club, in partnership with AI21 Labs, Entrée Capital, AWS, and NVIDIA.
The conversation? AI Agents—one of the most buzzed-about topics in AI today. Industry leaders and entrepreneurs shared insights, challenges, and opportunities in this rapidly evolving space.
A huge thank you to our incredible speakers and panelists!
@origoshen @StavleviNeumark @MickeHaslavsky @maypiamenta1 @BennyPorat @RonniZehavi @therealnirs @shargov
#ai #aiagents @EntreeCap -
AI21 Labs
@AI21Labs
3 events. 1 city. 🔥
Boston, you delivered.
✅ 80+ builders at the @sundai_club Hackathon
✅ 80+ AI devs at @AITinkerers meetup
✅ 20 @HarvardHBS founders building AI-first startups
We had so much fun building, brainstorming, and shipping projects with #maestro, our new agentic AI system!
Thanks to everyone who joined 🙌
#ai21 #maestro #hackathon #meetup #ai -
AI21 Labs
@AI21Labs
Today at @nvidia GTC! ⚡ Join AI21 Labs’ co-founder and co-CEO @origoshen as he takes the stage to discuss the transition from reasoning models to AI planning systems.
If you’re attending, let us know in the comments! -
AI21 Labs
@AI21Labs
Our CEO @origoshen on live stage at #humanx2025 presenting #ai21maestro AI21’s new approach to structured, multi-step AI execution. Finally, AI that follows and executes a plan. Thanks @stefanweitz and #humanx team for inviting us.
-
AI21 Labs reposted
AI21 Labs
@AI21Labs
Meet Maestro: @AI21Labs new AI Planning and Orchestration system. Unlike LLM-based agents, #maestro ensures reliable outcomes while optimizing latency and cost. This is AI that you can actually trust.
🔗 Sign up now! www.ai21.com/maestro
#ai21maestro #accuracyatscale #aiorchestrator #aiplanning -
AI21 Labs
@AI21Labs
Meet Maestro: @AI21Labs new AI Planning and Orchestration system. Unlike LLM-based agents, #maestro ensures reliable outcomes while optimizing latency and cost. This is AI that you can actually trust.
🔗 Sign up now! www.ai21.com/maestro
#ai21maestro #accuracyatscale #aiorchestrator #aiplanning -
AI21 Labs
@AI21Labs
Jamba Mini 1.6 outpacing Gemini 2.0 Flash, GPT-4o mini, and Mistral Small 3 on output speed.
artificialanalysis.ai/models/jamba-1-6-mini
This is what fast looks like! #buildopen #ai21jamba -
AI21 Labs reposted
AI21 Labs
@AI21Labs
Today we launched Jamba 1.6, the best open model for private enterprise deployment. AI21’s Jamba outperforms Cohere, Mistral and Llama on key benchmarks, including Arena Hard, and rivals leading closed models while maintaining unmatched speed and quality.
Now available on AI21’s Studio and @Hugging Face.
Learn more: www.ai21.com/jamba/ -
AI21 Labs
@AI21Labs
Today we launched Jamba 1.6, the best open model for private enterprise deployment. AI21’s Jamba outperforms Cohere, Mistral and Llama on key benchmarks, including Arena Hard, and rivals leading closed models while maintaining unmatched speed and quality.
Now available on AI21’s Studio and @Hugging Face.
Learn more: www.ai21.com/jamba/ -
AI21 Labs reposted
AI21 Labs
@AI21Labs
🛑 LLMs aren’t traditional software. You can’t unit-test AI—so how do you guarantee reliability?
LLMs ≠ 💾 #machinelearning #devtools -
AI21 Labs
@AI21Labs
Current AI systems are fundamentally broken.
Our co-founders and co-CEOs @origoshen and @yshoham share their thoughts on what the next phase of Enterprise AI will look like.
www.ai21.com/blog/accuracy-at-scale/ -
AI21 Labs
@AI21Labs
🛑 LLMs aren’t traditional software. You can’t unit-test AI—so how do you guarantee reliability?
LLMs ≠ 💾 #machinelearning #devtools -
AI21 Labs
@AI21Labs
Happy to share #yaap Yet Another AI Podcast - 1st episode #rag Evaluation - You're Doing It Wrong youtu.be/RsSkwpTmn8o via @YouTube
-
AI21 Labs
@AI21Labs
📣 Check out the latest from #amazonbedrock & AI21 Labs.
In this interview, our very own Chris Carpenter shares with @AWSAI's Brooke Jamieson how #jamba makes it easier for organizations to extract insights from long form documents, automate workflows and build faster, smarter AI solutions.
To hear more insights, listen and watch him here 👇
#aws #generativeai
AWS AI
@AWSAI
Learn about the latest from AI21 Labs & #amazonbedrock. 🤖
Jamba 1.5 is making it easier for businesses to extract insights from complex documents, automate workflows, & build faster, smarter #ai applications. #generativeai #aws
👉 go.aws/40EaucMThis media couldn't be displayed. See it on the original post. -
AI21 Labs
@AI21Labs
Excited to share that our paper, Jamba: Hybrid Transformer-Mamba Language Models, has been accepted to ICLR 2025!
We’re honored to contribute to this important platform and the collective knowledge of the AI community by sharing our work openly and submitting it for peer review.
This paper presents our Jamba family of open models built with a novel hybrid Transformer-Mamba architecture that combines the strengths of both approaches by delivering both efficiency and accuracy.
AI21’s Jamba has the #1 longest effective context window on the market per NVIDIA’s Ruler benchmark, making it the ideal choice for agentic workflows where long-context processing is mission critical.
Congrats to the co-authors and collaborators who contributed to this work - this paper represents a true team effort, and we’re proud to share it with the AI community. -
AI21 Labs
@AI21Labs
We’re excited to announce that our new course, ‘Build Long Context AI Apps with Jamba’, built in partnership with @AndrewYNg @DeepLearningAI is now live - and it’s currently free!
In this course, Chen Wang and Chen Almagor from AI21 Labs will walk you through how to use Jamba for:
✨ Long-context prompting
⚙️ Tool calling
📄 Analyzing long documents
🛠️ Conversational RAG
Plus, you’ll dive deep into the #jamba architecture and its unique design.
#jamballm #ai21labs #jamba
👉 Start learning today: www.deeplearning.ai/short-courses/build-long-context-ai-a… -
AI21 Labs
@AI21Labs
We are warning about scams being perpetrated by malicious actors falsely claiming to be AI21-related crypto/tokens. We explicitly clarify: AI21 has absolutely no connection, direct or indirect, to any cryptocurrency or tokens whatsoever. These cases have been reported to X and AI21 reserves the right to pursue legal measures against those involved.
Please stay vigilant and avoid any engagement with these scams.
#ai21 #scamalert -
AI21 Labs reposted
Amit Mandelbaum
@Amit_Mandelbaum
אם אתם מפתחים בעולם ה AI יש לנו יום מעניין מאוד בשבילכם, ואתם מוזמנים להירשם
אנחנו @AI21Labs ביחד עם @nvidia יצרנו יום מלא בהרצאות, שולחנות עגולים, ותמיכה טכנית שנוגעים בכל נושא בעולמות הפיתוח של GenAI. סוכנים , RAG, אימון מודלים, איך לחברות את ה LLM הנכון לכל משימה ועוד ועוד.
אוכל, פרסים שווים וסדנאות מעולות, והכל אצלנו במשרדים (המאוד יפים, אבל אני לא אובייקטיבי), מוזמנים להירשם (מספר המקומות מוגבל):
www.ai21.com/lp/ai21-nvidia-ai-devcon -
AI21 Labs
@AI21Labs
Come join us at #sftechweek! Sign up now: lu.ma/nho5h281
Katia Ameri
@KatiaAmeri
🎉 Big news! The official calendar of events for #sftechweek and #latechweek is now LIVE!
And thrilled to announce that this is officially our BIGGEST Tech Week, ever!
- 850+ events 🤯
- San Francisco: Oct 7-13 🌉
- Los Angeles: Oct 14-20 🌴
Link to register below to see all events👇 -
AI21 Labs reposted
Ishaan
@ishaan_jaff
🔥 Today we're launching support for @AI21Labs Jamba 1.5-mini, Jamba 1.5-large on @LiteLLM
Use 100+ LLMs in unified interface with LiteLLM Gateway or LiteLLM Python SDK
👉 Start here: docs.litellm.ai/docs/providers/ai21
🔥 Use AI21 models on Vertex AI here: docs.litellm.ai/docs/providers/vertex -
AI21 Labs
@AI21Labs
We know that Jamba 1.5 models are the fastest, but the question is - how fast? @ArtificialAnlys tested our models to find out 😎The image below shows the throughput for various models (with prompt length = 10K tokens). Jamba 1.5 models are a whole lot faster – and that speed delta only grows with longer prompts. >>
Post not available here, view on X
AI21 Labs
@AI21Labs
But our cost, efficiency and speed don't come at the expense of quality. In the following chart, Jamba 1.5 Large and Mini both show a great balance between speed and quality (QI is where you want to be). >>
AI21 Labs
@AI21Labs
Want to test it yourself? Both models are available at 🤗 Check them out: huggingface.co/collections/ai21labs/jamba-15-66c44befa474…
-
AI21 Labs
@AI21Labs
Introducing the Jamba 1.5 Model Family on @googlecloud's Vertex AI:
- Simplify development and evaluation with advanced tools and intuitive API calls.
- Focus on innovation with fully managed infrastructure and cost-effective, pay-as-you-go pricing.
- Ensure data security with robust privacy controls and compliance certifications.
cloud.google.com/blog/products/ai-machine-learning/jamba-…
#vertexai #googlecloud #googlecloudpartner -
AI21 Labs
@AI21Labs
Our team had a blast at the @agihouse_org agents hackathon yesterday in the Bay Area with @langchain, @GroqInc , @DeepLearningAI & @AndrewYNg.
Our new Jamba 1.5 models are optimized for agent creation with a ton of built in features like:
- Function calling
- Structured JSON output
- Digesting doc objects
- Generating citations -
AI21 Labs reposted
Ofer Mendelevitch
@ofermend
Exciting to see this new model from @AI21Labs, using Mamba and having such great performance. Just added jamba-1.5-mini to @vectara Hallucination Leaderboard, and it show 2.9% hallucination rate - quite impressive.
huggingface.co/spaces/vectara/leaderboard -
AI21 Labs reposted
xundecidability
@xundecidability
🏇 giddyup! exciting new jamba models (mamba+transformers =) from @AI21Labs landed in @openrouter 🚂🔜🚀
-
AI21 Labs
@AI21Labs
📄Jamba-1.5 whitepaper is out!
The whitepaper details the architecture, training schemes, novelties and in-depth evaluations of our new long context hybrid SSM-Transformer models - Jamba-1.5-Large and Jamba-1.5-Mini.
Arxiv: arxiv.org/abs/2408.12570
Here are some highlights and insights from the paper 👇1/7Post not available here, view on X
AI21 Labs
@AI21Labs
Jamba-1.5 models perform well in multiple languages, even though we include only a very small fraction of non-english data in the post-training phase. Therefore, we speculate the models are able to use the learned multilingual capabilities from the pre-training phase. 6/7
AI21 Labs
@AI21Labs
We found our efficient Jamba architecture to be advantageous in long context fine-tuning, as it allows for greater speed and lower cost. Therefore, we could experiment with multiple different training recipes during the fine-tuning phase. This is especially interesting for all the practitioners out there, since our models are released with open weights: huggingface.co/collections/ai21labs/jamba-15-66c44befa474…. We can’t wait to see what the community will build on this technology 🤗 7/7
-
AI21 Labs reposted
Haider.
@haider1
Jamba 1.5 is out
Here is some important information + live code test:
» Improve document summarization and analysis with a 256K context window
» Up to 2.5X faster on long contexts
» Outperforms larger models like Llama 3.1 70B and 405B
» Supports English, Spanish, French, and more
» Supports JSON, function calls, and document objects
AI21 Labs
@AI21Labs
We released the #jamba 1.5 open model family:
- 256K #contextwindow
- Up to 2.5X faster on #longcontext in its size class
- Native support for structured JSON output, function calling, digesting doc objects & generating citations
twtr.to/giIEE
#ai #llm #ai21jamba -
AI21 Labs reposted
Tri Dao
@tri_dao
Hybrid Mamba / Transformer architecture works really well, especially for long context and fast inference
AI21 Labs
@AI21Labs
We released the #jamba 1.5 open model family:
- 256K #contextwindow
- Up to 2.5X faster on #longcontext in its size class
- Native support for structured JSON output, function calling, digesting doc objects & generating citations
twtr.to/giIEE
#ai #llm #ai21jamba -
AI21 Labs
@AI21Labs
We wanted to share some more granular details about the Jamba 1.5 model family - and specific benchmarks on latency, context window, and quality. [1/6]
Post not available here, view on X
AI21 Labs
@AI21Labs
Both Jamba 1.5 models are faster than competitors of a similar size, with up to 2.5X faster inference on long contexts. [3/6]
AI21 Labs
@AI21Labs
Jamba’s speed also shows in OTPS, as seen in @ArtificialAnlys. Jamba 1.5 Mini ranks the fastest @ 10K contexts [4/6]
AI21 Labs
@AI21Labs
On Arena Hard benchmark, Jamba 1.5 Mini scores an excellent 46.1, surpassing even larger models. Jamba 1.5 Large scores 65.4, above Llama 3.1 70B and Llama 3.1 405B. [5/6]
AI21 Labs
@AI21Labs
To make sure Jamba 1.5 Large fits a single node, we developed ExpertsInt8, a quantization technique for MoE models. We are now able to fit the model with its entire 256K context window on a single node, without losing quality or having a complicated quantization process. [6/6]
-
AI21 Labs reposted
NVIDIA AI Developer
@NVIDIAAIDev
🔍 Explore @AI21Labs' Jamba 1.5, designed for diverse AI tasks like content creation and data insights.
➡️ nvda.ws/3SXyKU0
Using transformer and Mamba architectures, this MoE model ensures top efficiency and extensive context handling. Experience Jamba 1.5 API as NVIDIA NIM microservice from the NVIDIA API catalog. -
AI21 Labs reposted
Google Cloud Tech
@GoogleCloudTech
📣 Two new models have arrived in Vertex AI Model Garden: Jamba 1.5 Mini and Jamba 1.5 Large from @AI21Labs!
Both models in the Jamba 1.5 Model Family feature a 256K context window + Mamba-Transformer architecture for speed and long context efficiency → goo.gle/4dKxyvo -
AI21 Labs reposted
Yam Peleg
@Yampeleg
Final thoughts:
I really think this model is crazy.
The first mamba-hybrid being able to compete with the top performers.
While it takes time to validate these performance, you for sure couldn't get this type of performance at these prices before. period. -
AI21 Labs
@AI21Labs
Thanks @togethercompute - we're exciting to be making our Jamba models available on Together AI soon!
Together AI
@togethercompute
Massive congratulations to our partners at @AI21labs on the release of their Jamba 1.5 models.
Jamba 1.5 is a game-changer for long context and RAG:
- 256K context window
- 2.5x faster on long contexts
- JSON output, function calling, citation mode
Open innovation at its finest. Coming soon on the together.ai platform.
👏Incredible work from @AI21Labs
AI21 Labs
@AI21Labs
We released the #jamba 1.5 open model family:
- 256K #contextwindow
- Up to 2.5X faster on #longcontext in its size class
- Native support for structured JSON output, function calling, digesting doc objects & generating citations
twtr.to/giIEE
#ai #llm #ai21jamba -
AI21 Labs reposted
Yam Peleg
@Yampeleg
Thursday AI bombs started early today!!
Jamba-1.5 just dropped!
Largest mamba to date.
(mamba + attention + moe hybrid)
TL;DR: long context, low vram, GREAT scores!
𝙹𝚊𝚖𝚋𝚊-𝟷.𝟻-𝙼𝚒𝚗𝚒 (𝟷𝟸𝙱/𝟻𝟸𝙱):
huggingface.co/ai21labs/AI21-Jamba-1.5-Mini
𝙹𝚊𝚖𝚋𝚊 𝟷.𝟻 𝙻𝚊𝚛𝚐𝚎 (𝟿𝟺𝙱/𝟹𝟿𝟾𝙱):
huggingface.co/ai21labs/AI21-Jamba-1.5-Large -
AI21 Labs reposted
Artificial Analysis
@ArtificialAnlys
At 10k input tokens and beyond, our independent performance benchmarks show the benefits of Jamba’s state space/transformer hybrid architecture for keeping TTFT low and output speed high.
-
AI21 Labs reposted
Artificial Analysis
@ArtificialAnlys
@AI21Labs launches Jamba 1.5 Mini and Large! These models utilize a hybrid state space/transformer architecture to maintain output speed with long context inputs⚡
Both Mini and Large launch with 256k context windows, the largest of all major open weights models.
Their hybrid state space/transformer architecture supports the models in maintaining high performance over long prompt lengths. Jamba 1.5 is leading in speed compared to models of a similar intelligence class (comparing to the median performance across providers for each model).
Artificial Analysis has independently benchmarked Jamba 1.5 across quality, API performance and price.
See below for further charts and a link to our full analysis of AI21’s models: -
AI21 Labs
@AI21Labs
We released the #jamba 1.5 open model family:
- 256K #contextwindow
- Up to 2.5X faster on #longcontext in its size class
- Native support for structured JSON output, function calling, digesting doc objects & generating citations
twtr.to/giIEE
#ai #llm #ai21jamba -
AI21 Labs reposted
Philipp Schmid
@_philschmid
New Jamba 1.5 release! @AI21Labs updated their production-scale Mamba implementation! Jamba is a hybrid SSM-Transformer MoE rivaling other open LLMs. Jamba 1.5 comes into sizes mini 52B-A12B and large 398B-A94B 🚀
TL;DR:
2️⃣ Jamba 1.5 comes as Mini (12B active/52B total) and Large (94B active/398B total)
📏 Supports 256K context & optimized for long context RAG
🌍 Supports 8 languages: English, Spanish, French, Portuguese, Italian, Dutch, German, Arabic, and Hebrew
💼 Optimized for function calling and structured output (JSON)
🧪 New ExpertsInt8 quantization technique for more efficient deployment
⚡ Up to 2.5X faster inference in long context compared to Decoder-only models.
💻 Mini should fit up to 140K context on a single A100 80GB
🔓 Custom license with limitations at > $50M revenue
🤗 Available on @huggingface and in Transformers
🏆 Rivals open LLMs on Open LLM Leaderboard Benchmarks
❌ No information about the training process or data
Blog: www.ai21.com/blog/announcing-jamba-model-family
Jamba 1.5 mini: huggingface.co/ai21labs/AI21-Jamba-1.5-Large
Jamba 1.5 large: huggingface.co/ai21labs/AI21-Jamba-1.5-Large -
AI21 Labs
@AI21Labs
Last chance to sign up for the #open-Source #agent #hackathon with @AndrewYNg at the @agihouse_org!
Hack with the movers-and-doers of Gen-AI from @DeepLearningAI, @GroqInc, @langchain, and of course, some of the builders of #jamba from @AI21Labs
🔗 www.agihouse.org/events/open-source-agent-hackathon-20240… -
AI21 Labs
@AI21Labs
Hack by popular demand!
Open-Source Agent #hackathon with @AndrewYNg & @DeepLearningAI at @agihouse_org Saturday, *August 24th*
Stellar schedule planned w/ secret guest speakers, alongside co-hosts, @GroqInc & @langchain.
Reserve your spot now --> www.agihouse.org/events/open-source-agent-hackathon-20240… -
AI21 Labs
@AI21Labs
Jamba support is now live on vLLM 🎉
Due to its novel hybrid SSM-Transformer arch, Jamba didn’t work out-of-the-box in vLLM. our own @MorZusman worked together with @vllm_project to integrate Jamba for an efficient serving in vLLM 🙌
github.com/vllm-project/vllm/releases/tag/v0.5.1
AI21 Labs
@AI21Labs
Things in this PR:
- Jamba modeling file, supporting Jamba variations
- Mamba cache management (also benefit other Mamba-based models)
- Send more request related properties to the forward pass, for model-specific implementations (including things needed for speculative decoding)
AI21 Labs
@AI21Labs
BTW, Since Jamba supports a 256K context with high throughput, we also stumbled upon an issue where the fused_moe kernel didn’t work well in long contexts. Others seems to have had this too, according to some other open issues
AI21 Labs
@AI21Labs
@avshalomm solved it by utilizing the fact that actually, there is no interaction between different tokens in the MoE block, so we can iterate over the long context in chunks. 😎
This was also merged and now fixed in vLLM -
AI21 Labs
@AI21Labs
On Saturday, we hosted our first Bay Area hackathon at @agihouse_org, inviting devs to build apps with #jamba.
After 8 hrs of intense coding, each team presented a live demo leveraging Jamba’s massive 256K context window.
👇 Here are the highlights
AI21 Labs
@AI21Labs
🥇In 1st: FBI AGI
A real-time incrimination tool against predators on social media.
Built by Alex Sima & Apurva Mishra
AI21 Labs
@AI21Labs
🥈In 2nd: Excel Mamba
Complex financial Excel formula-generation using AI agents.
Built by Aniket Shirke, Bhavya Bahl, & Rohit Jena
AI21 Labs
@AI21Labs
🥉In 3rd: AI Pals
The team built avatars that could communicate via voice-over. By leveraging Jamba’s long context, entire books were used within a single prompt, thereby decreasing latency.
Built by Anto Patrex, Anand Vallamsetla, Justin Jung, & Lars Traaholt Vågnes -
AI21 Labs reposted
OpenRouter
@OpenRouter
Warm welcome to @AI21Labs: Jamba Instruct is now on OpenRouter! 🐍
1.8M tokens processed today. 256k context. Hybrid transformer architecture based on Mamba. Link below!
AI21 Labs
@AI21Labs
We just released Jamba-Instruct!
Built from our groundbreaking SSM-Transformer Jamba architecture, Jamba-Instruct brings the same technological innovation to the enterprise via an aligned model.
With leading quality benchmarks, a 256K context window, and the most competitive pricing in its size class, you’re getting the most value for your money.
Read more in our blog: www.ai21.com/blog/announcing-jamba-instruct
(1/4) -
AI21 Labs
@AI21Labs
Live from the @agihouse_org it’s the first ever Jamba hackathon! What long context use cases would you want to build with Jamba’s 256k context window?
-
AI21 Labs
@AI21Labs
#jamba set new standards for LLMs, but this SSM-Transformer's just getting started.
Come hack with us @ the AGI House to further redefine the capabilities of genAI.
Expect prizes, including $5K credits to build apps w/ Jamba & exclusive swag.
Sign up: shorturl.at/v0f4T?hss_channel=tw-1166332664569368576 -
AI21 Labs
@AI21Labs
We’re thrilled to announce that Jamba-Instruct will soon be launching on @Snowflake Cortex AI!
“With AI21's groundbreaking Jamba-Instruct offered in Cortex AI, our customers will now be able to seamlessly connect their data to build transformative GenAI applications with AI21's powerful models.”
- Baris Gultekin, Head of AI at @Snowflake
Blog here: twtr.to/whZN_
#snowflakecortex #jamba #genai #llms -
AI21 Labs
@AI21Labs
We just released Jamba-Instruct!
Built from our groundbreaking SSM-Transformer Jamba architecture, Jamba-Instruct brings the same technological innovation to the enterprise via an aligned model.
With leading quality benchmarks, a 256K context window, and the most competitive pricing in its size class, you’re getting the most value for your money.
Read more in our blog: www.ai21.com/blog/announcing-jamba-instruct
(1/4)Post not available here, view on X
AI21 Labs
@AI21Labs
(3/4) Jamba-Instruct results on long-context QA benchmarks, conducted using the same method outlined in section 5.2.2 of our Jamba base model whitepaper.
AI21 Labs
@AI21Labs
(4/4) With a 256K context window, Jamba-Instruct boasts the largest context window in its size class.
-
AI21 Labs reposted
NYSE 🏛
@NYSE
.@origoshen, Co-Founder + Co-CEO of @AI21Labs, talks about growth opportunities for the company following a $208 million Series C funding round, and shares his perspective on the future of AI on #nysefloortalk with @JudyKShaw
This media couldn't be displayed. See it on the original post. -
AI21 Labs
@AI21Labs
Building a RAG solution is easy. Building a great one is not.
In our guest blog on @streamlit, our team explores the intricacies of how AI21's Contextual Answers Task-Specific Model & our RAG Engine generate context-based answers grounded in your proprietary organizational data.
You’ll find step-by-step instructions on how to effortlessly build a Multi-Doc Q&A app powered by our GenAI solutions.
Check it out here:
blog.streamlit.io/ai21_grounded_multi_doc_q-a/ -
AI21 Labs
@AI21Labs
Introducing Jamba, our groundbreaking SSM-Transformer open model!
As the first production-grade model based on Mamba architecture, Jamba achieves an unprecedented 3X throughput and fits 140K context on a single GPU.
🥂Meet Jamba www.ai21.com/jamba
🔨Build on @huggingface
AI21 Labs
@AI21Labs
🥂Meet Jamba
www.ai21.com/jamba
🔨Build with Jamba
huggingface.co/ai21labs/Jamba-v0.1
📝Read more
www.ai21.com/blog/announcing-jamba
#jamba #mamba #transformer #aiarchitecture #ssm
AI21 Labs
@AI21Labs
📄The Jamba whitepaper details our in-depth ablations on this novel hybrid SSM-Transformer architecture, and how we chose to interleave Mamba, Transformer and MoE. arxiv.org/abs/2403.19887
-
AI21 Labs reposted
meng shao
@shao__meng
Paper of Jamba: A Hybrid Transformer-Mamba Language Model
Jamba, a novel architecture which combines Attention and Mamba layers, with MoE modules, and an open implementation of it, reaching state-of-the-art performance and supporting long contexts.
We showed how Jamba provides flexibility for balancing performance and memory requirements, while maintaining a high throughput. We experimented with several design choices such as the ratio of Attention-to-Mamba layers and discussed some discoveries made during the development process, which will inform future work on hybrid attention–state-space models. To
facilitate such research, we plan to release model checkpoints from smaller-scale training runs.
The largest model we provide with this release has 12B active and 52B total available parameters, supporting context lengths of up to 256K tokens and fitting in a single 80GB GPU even when processing 140K-token texts.
arxiv.org/pdf/2403.19887.pdf
AK
@_akhaliq
Jamba
A Hybrid Transformer-Mamba Language Model
present Jamba, a new base large language model based on a novel hybrid Transformer-Mamba mixture-of-experts (MoE) architecture. Specifically, Jamba interleaves blocks of Transformer and Mamba layers, enjoying the benefits of -
AI21 Labs reposted
AK
@_akhaliq
Jamba
A Hybrid Transformer-Mamba Language Model
present Jamba, a new base large language model based on a novel hybrid Transformer-Mamba mixture-of-experts (MoE) architecture. Specifically, Jamba interleaves blocks of Transformer and Mamba layers, enjoying the benefits of -
AI21 Labs reposted
Philipp Schmid
@_philschmid
Last Week @AI21Labs released the production-scale Mamba implementation, and today, they released their paper. 🧐
Jamba introduces a new hybrid Transformer-Mamba mixture-of-experts architecture offering state-of-the-art performance but with significant improvements on long context. 🐍
Jamba Architecture
1️⃣ Introduces Jamba blocks. A combination of Mamba layers and Transformer, with mixture-of-experts (MoE) modules in between.
Jamba 52B uses Jamba blocks with 8 layers and a 1:7 Attention-to-Mamba ratio with MoE layers in between. MoE Layers have 16 experts with 2 experts at each token.
2️⃣ Jamba 52B has 4 Jamba blocks with a total of 4 Transformer and 28 Mamba blocks.
3️⃣ For Transformer blocks grouped-query attention (GQA), SwiGLU activation function were used.
4️⃣ Added RMSNorm to stabilize Mamba layers at large scale to prevent loss spikes.
5️⃣ No explicit positional information (e.g., RoPE), as Mamba layers may provide implicit position information.
6️⃣ Mixture-of-experts improves the Attention-Mamba hybrid, no details on why, it's just better based on experiments
Additional details
🤏🏻 Fits in a single 80GB GPU in 8bits with 140k context
⚡️ Trained on NVIDIA H100 using FSDP with Tensor Parallelism, Sequence Parallelism, Expert Parallelism
📅 Text Data from the Web, books, and code, with the last update in March 2024, no more details
🏆 Matches performance of Mixtral 8x7B and outperforms llama 70B
🧠 Attention-to-Mamba 1:3 or 1:7 with virtually no performance difference
❌ Paper does not specify the dataset size or number of training tokens -
AI21 Labs
@AI21Labs
📄Jamba whitepaper is out!
The whitepaper details our in-depth ablations on this novel hybrid SSM-Transformer architecture, and how we chose to interleave Mamba, Transformer and MoE.
arxiv.org/abs/2403.19887
Here are some highlights from the paper 👇1/6Post not available here, view on X
AI21 Labs
@AI21Labs
Combining Transformer, Mamba & MoE allows flexibility in balancing low memory usage, high throughput, and high quality. Jamba’s KV cache – which becomes a limiting factor when scaling context in pure Transformers – is 8x smaller compared to a pure Transformer. 5/6
AI21 Labs
@AI21Labs
We found that scaling Mamba to substantial sizes isn’t trivial and causes unstable training. We had to make some adjustments to the Mamba layers to stabilize them, as can be seen in the graph below (training loss before & after the change) 6/6
-
AI21 Labs reposted
Philipp Schmid
@_philschmid
Yesterday AI21 Labs released Jamba, the first production-scale Mamba implementation as a hybrid SSM-Transformer MoE 🐍 And today, you can already finetune it with Hugging Face TRL.
@Dorialexander shared a working Qlora script using 4-bit quantization on an A100 GPU (for now, disabled fast Mamba kernels). 🚀
Check out his repo for the details: github.com/Pleias/Various-Finetuning/blob/main/finetuning… -
AI21 Labs reposted
1LittleCoder💻
@1littlecoder
🥹 Jamba is truly amazing!
Everyone speaks about Long Context. But it's been mostly useful for ingesting in-context learning.
But Jamba seems to be the first Model offering a great throughput even for higher context! -
AI21 Labs reposted
Maxime Labonne
@maximelabonne
I played a little with Jamba: it looks like an amazing model.
In terms of architecture, the MoE implementation is very close to Mixtral's. What's great about it is that it hasn't been fine-tuned. Curious to see how much improvement we can get through SFT.
I made a little notebook if you want to play with it (I used an A100 80GB instead of Colab though): colab.research.google.com/drive/1swHmnAyixlPBA37zA1eDxrTt…
AI21 Labs
@AI21Labs
Introducing Jamba, our groundbreaking SSM-Transformer open model!
As the first production-grade model based on Mamba architecture, Jamba achieves an unprecedented 3X throughput and fits 140K context on a single GPU.
🥂Meet Jamba www.ai21.com/jamba
🔨Build on @huggingface -
AI21 Labs reposted
Aleksa Gordić (水平问题)
@gordic_aleksa
Extremely cool new model release from @AI21Labs - Jamba - and it's not even a transformer! It's a hybrid model that combines Mamba (structured state space model), transformer layers and MoE technique, and it's a first production-grade Mamba based model!
* It's a 52B MoE with 12B params active during inference.
* Due to it being Mamba-based it has a much better throughput and efficiency.
* 3x throughput on long contexts compared to Mixtral.
* It has a 256k context window! Out of those 140k context can be achieved on a single GPU which makes it much easier to deploy.
* Perf-wise it looks to be on-par with Mixtral.
* It's open-weight model (licensed under Apache 2.0)
Really exciting, this is apparently the first time someone scaled Mamba upward from 3B. 52B it is. :)
HuggingFace: huggingface.co/ai21labs/Jamba-v0.1
Blog: www.ai21.com/blog/announcing-jamba -
AI21 Labs reposted
swyx 🔜 aie.codes
@swyx
incredibly impressed by @AI21Labs' Jamba today. This is the first legitimate Mixtral-killer we've seen and it came out of "nowhere":
buttondown.email/ainews/archive/ainews-jamba-mixture-of-a…
They've helped me redefine my idea of a model "weight class" from "number of parameters" (increasingly outdated with MoEs and mixed arches) to "default system requirements". It ~matches Mixtral on every benchmark, but every architecture choice seems ideal to max out a single 80GB A100 GPU, which makes it WAY better (like, 2-3x better in every efficiency/speed dimension you care about) for long context usecases.
This efficient long context stuff is what Mamba promised but it wasn't as clear until now and has again shifted my timelines on when/if Mamba dethrones Transformers (more likely they mix layers just like Jamba and StripedHyena show). Also dont miss Cobra - x.com/swyx/status/1773431296687247406 Mamba may also just crush in multimodal applications as well.
Lastly - i think releasing only base model is kinda genius. Jamba has PEFT support OOTB. AI21 is asking you not to judge them on their instruct tuning - surely armies of cracked eastern european alpha coders will do that for them - and they surely will given the very strong base metrics.
Let a thousand Jambas bloom.
What a time to be alive.
swyx 🔜 aie.codes
@swyx
TIL someone extended Mamba models to vision and just after 1 day of training it destroys LLaVA et al on everything except OCR
this state space model thing might have something to it -
AI21 Labs reposted
Carlos E. Perez
@IntuitMachine
OMG! Jamba! 3X throughput and fits 140K context on a single GPU!
-
AI21 Labs
@AI21Labs
Find our new model on @huggingface.
Philipp Schmid
@_philschmid
Jamba released! @AI21Labs just released the first production-scale Mamba implementation! Jamba is a hybrid SSM-Transformer MoE rivaling open transformer-based LLMs 🚀
TL;DR:
🧠 52B parameters with 12B active during generation
👨🏫 16 experts with 2 active in generation
🆕 New architecture with Joint Attention and Mamba
⚡️ Supports 256K context length
💻 Fits up to 140K context on a single A100 80GB
🚀 3X throughput on long contexts compared to Mixtral 8x7B
🔓 Released under Apache 2.0
🤗 Available on @huggingface & Transformers (>4.38.2)
🏆 Rivals open LLMs on Open LLM Leaderboard Benchmarks
❌ No information about training data or language support
Blog: www.ai21.com/blog/announcing-jamba
Model: huggingface.co/ai21labs/Jamba-v0.1 -
AI21 Labs
@AI21Labs
Thanks @tri_dao! We enjoyed the opportunity to extend and build on top of Mamba. The Jamba white paper is coming soon and has some ablations we will share on this subject.
Tri Dao
@tri_dao
Wow this is a big deal, first large-scale Mamba-based model! Mamba layers brings much longer context and higher inference throughput. Having 4 attention layers seem to be the sweet spot to get the best of Transformer & Mamba architectures.
AI21 Labs
@AI21Labs
Introducing Jamba, our groundbreaking SSM-Transformer open model!
As the first production-grade model based on Mamba architecture, Jamba achieves an unprecedented 3X throughput and fits 140K context on a single GPU.
🥂Meet Jamba www.ai21.com/jamba
🔨Build on @huggingface -
AI21 Labs reposted
Niels Rogge
@NielsRogge
What a week, huh?
Captain, it's only Thursday
- @databricks releases DBRX, a 132B MoE
- @AI21Labs releases Jamba, a 52B Transformer/Mamba MoE
- @AlibabaGroup releases Qwen-1.5-MoE, a 2.7B MoE
All on hf.co! 🤗
Learn about MoE here: huggingface.co/blog/moe -
AI21 Labs reposted
Mark Tenenholtz
@marktenenholtz
First scaled-up Mamba release is here from AI21!
Jamba is a hybrid SSM/Transformer MoE with 52B parameters (12B active), 256k context length (140k fits on a single A100), and significantly better inference throughput.
The benchmarks make sense: looks like it outperforms transformers on tasks with generative reasoning, but loses out on multiple choice type tasks where attention-focused architectures tend to thrive.
It's not clear how many tokens it was trained on. What we need to see next is a lab releasing similar-scale transformers/mamba/jamba type architectures. It's hard to tell how much of the performance is architecture (besides the obvious inference benefits), and how much is data mixture.
My guess is that it's 99% data mixture. We may now be seeing a diverging path, where fully attention-based architectures are better for extraction-type tasks, and Mamba-type models are better for other types. -
AI21 Labs
@AI21Labs
And 140K context on a single GPU. @ArtificialAnlys We look forward to seeing those benchmarks!
Artificial Analysis
@ArtificialAnlys
Jamba model launch takeaways - Potentially a new leader for ultra-long prompt use-cases (RAG)
‣ First open-source model of this size to combine MAMBA state-space model architecture, Mixture-Of-Experts (MOE) and the transformer
‣ 256k context window, more than 2X the size of the next largest open-source model (Code Llama 70B's 100k) we measure
‣ High expected throughput tokens/s as with its MOE architecture, 12B of its 52B parameters are active at inference. For shorter prompts, expect faster than Grok-1 & Llama 2 but slower than Mixtral 8x7B
‣ Presents a potentially very attractive offer for long input token prompts / RAG as throughput scales with input token size due to MAMBA architecture. A21 declare 3X Mixtral 8x7B's tokens/s at 128k context window lengths
Congratulations @AI21Labs. We look forward to benchmarking these declared speeds, particularly over long input token lengths 👀 -
AI21 Labs reposted
Alexander Doria
@Dorialexander
Open LLMs strike back. Not only a strong model but a disruptive architecture you won't find in closed AI.
AI21 Labs
@AI21Labs
Introducing Jamba, our groundbreaking SSM-Transformer open model!
As the first production-grade model based on Mamba architecture, Jamba achieves an unprecedented 3X throughput and fits 140K context on a single GPU.
🥂Meet Jamba www.ai21.com/jamba
🔨Build on @huggingface -
AI21 Labs reposted
Carter Leffen
@carterleffen
@AI21Labs bringing the heat with a hybrid architecture! can’t wait to see the fine-tunes
AK
@_akhaliq
AI21 Labs presents Jamba
SSM-Transformer open model
production-grade model based on Mamba architecture, Jamba achieves an unprecedented 3X throughput and fits 140K context on a single GPU. -
AI21 Labs reposted
Rohan Paul
@rohanpaul_ai
Jamba, the world’s first production-grade Mamba based model is released by @AI21Labs
📌 New architecture with Joint Attention and Mamba
📌 Supports 256K context length
📌The only model in its size class that fits up to 140K context on a single GPU
📌3X throughput on long contexts compared to Mixtral 8x7B -
AI21 Labs reposted
Casper Hansen
@casper_hansen_
This year is going to be amazing:
256K context total, 140K context in ONE GPU????
Close performance to Mixtral while having 3x higher throughput. Apache 2.0. What a release!
AI21 Labs
@AI21Labs
Introducing Jamba, our groundbreaking SSM-Transformer open model!
As the first production-grade model based on Mamba architecture, Jamba achieves an unprecedented 3X throughput and fits 140K context on a single GPU.
🥂Meet Jamba www.ai21.com/jamba
🔨Build on @huggingface -
AI21 Labs reposted
Simon Willison
@simonw
This one looks like a big deal. It's not a pure transformer model - it's a combination of both Mamba and transformer, which appears to address different limitations present in each architecture
AI21 Labs
@AI21Labs
Introducing Jamba, our groundbreaking SSM-Transformer open model!
As the first production-grade model based on Mamba architecture, Jamba achieves an unprecedented 3X throughput and fits 140K context on a single GPU.
🥂Meet Jamba www.ai21.com/jamba
🔨Build on @huggingface -
AI21 Labs
@AI21Labs
RT @danielmerja: Wake up honey, new LLM on Mamba architecture just dropped
AI21 Labs
@AI21Labs
Introducing Jamba, our groundbreaking SSM-Transformer open model!
As the first production-grade model based on Mamba architecture, Jamba achieves an unprecedented 3X throughput and fits 140K context on a single GPU.
🥂Meet Jamba www.ai21.com/jamba
🔨Build on @huggingface -
AI21 Labs reposted
Yam Peleg
@Yampeleg
BREAKING!
AI21 just dropped an open source Mamba!
- MoE with 52B parameters.
- Active parameters: 12B.
- 256K Context length.
- Competitive performance to Mixtral!
- Open weights: Apache 2.0.
[I am visiting them in an hour so i'll get u some info!]
www.ai21.com/blog/announcing-jamba -
AI21 Labs reposted
Alex Volkov
@altryne
whoah ok, new architecture models are dropping on #thursdai
twitter.com/i/spaces/1OwxWYAapBVGQ
Jamba is a combined Mamba + Joint Attention 256K context support SSM-Transformer MoE (that's a mouthful) from @AI21Labs 🔥
See Phillip's announcement for all details and come chat with us about it in 30 min
Philipp Schmid
@_philschmid
Jamba released! @AI21Labs just released the first production-scale Mamba implementation! Jamba is a hybrid SSM-Transformer MoE rivaling open transformer-based LLMs 🚀
TL;DR:
🧠 52B parameters with 12B active during generation
👨🏫 16 experts with 2 active in generation
🆕 New architecture with Joint Attention and Mamba
⚡️ Supports 256K context length
💻 Fits up to 140K context on a single A100 80GB
🚀 3X throughput on long contexts compared to Mixtral 8x7B
🔓 Released under Apache 2.0
🤗 Available on @huggingface & Transformers (>4.38.2)
🏆 Rivals open LLMs on Open LLM Leaderboard Benchmarks
❌ No information about training data or language support
Blog: www.ai21.com/blog/announcing-jamba
Model: huggingface.co/ai21labs/Jamba-v0.1 -
AI21 Labs reposted
meng shao
@shao__meng
Introducing Jamba of @AI21Labs :
AI21's Groundbreaking SSM-Transformer Model
Debuting the first production-grade Mamba-based model delivering best-in-class quality and performance.
Key Features:
🎉First production-grade Mamba based model built on a novel SSM-Transformer hybrid architecture
🎉3X throughput on long contexts compared to Mixtral 8x7B
🎉Democratizes access to a massive 256K context window
🎉The only model in its size class that fits up to 140K context on a single GPU
🎉Released with open weights under Apache 2.0
🎉Available on Hugging Face and coming soon to the NVIDIA API catalog
AI21 Labs
@AI21Labs
Introducing Jamba, our groundbreaking SSM-Transformer open model!
As the first production-grade model based on Mamba architecture, Jamba achieves an unprecedented 3X throughput and fits 140K context on a single GPU.
🥂Meet Jamba www.ai21.com/jamba
🔨Build on @huggingface -
AI21 Labs reposted
Prashant
@Prashant_1722
A new AI model that can handle more context
Meet Jamba, groundbreaking SSM-Transformer open model from @AI21Labs
As the first production-grade model based on Mamba architecture, Jamba achieves an unprecedented 3X throughput and fits 140K context on a single GPU.
🥂Meet Jamba ai21.com/jamba
🔨Build on @huggingface
#ai #machinelearning #gpt -
AI21 Labs reposted
AK
@_akhaliq
AI21 Labs presents Jamba
SSM-Transformer open model
production-grade model based on Mamba architecture, Jamba achieves an unprecedented 3X throughput and fits 140K context on a single GPU. -
AI21 Labs reposted
TechCrunch
@TechCrunch
AI21 Labs’ new AI model can handle more context than most tcrn.ch/4cDMKL9
-
AI21 Labs
@AI21Labs
Stop struggling, start scaling.📈
In our upcoming webinar, we'll be sharing key strategies to successfully deploy GenAI at scale, including:
✨ Generating reliable outputs
✨ Methods for mitigating risks
✨ Integrating custom models
Secure your spot: aib-resources.aibusiness.com/free/w_defa5858/ -
AI21 Labs
@AI21Labs
Coming to #himss24? Learn how our #genai solutions deliver unmatched accuracy and privacy-sensitive responses. 🔐
🗓️ Schedule a 1:1 with us at the event: www.ai21.com/meeting/contact -
AI21 Labs
@AI21Labs
We've just rolled out some powerful upgrades to our @langchain integration, incorporating our Jurassic-2 models, alongside chat and embedding features! ⚙️
With this partnership, AI21's models are now more accessible, enabling developers to build impactful enterprise solutions.
LangChain
@LangChain
@AI21Labs Foundation and Task-Specific Models 🎯
Unlock the full potential of AI21's Jurassic-2 and Task-Specific models directly in LangChain! Effortlessly build and scale generative AI applications that meet today's business needs. No NLP expertise required; just seamless integration for text generation and comprehension. Get started now and transform your projects with the power of AI21 and LangChain.
python.langchain.com/docs/integrations/chat/ai21 -
AI21 Labs
@AI21Labs
Myth: General-purpose #llms can solve all enterprise business challenges.
Fact: Smaller, specialized models, deliver superior results.
Learn more about our Task-Specific models. 👇
www.ai21.com/meeting/contact -
AI21 Labs
@AI21Labs
We're here at #nrf2024! Swing by booth #6020, chat with our team, and discover how retailers are benefiting from #genai : ai21-labs.webflow.io/studio/national-retail-federation-co…
-
AI21 Labs
@AI21Labs
🛒Major retailers use #genai for personalized experiences.
Explore how the industry leverages AI solutions for customization in our latest blog: www.ai21.com/blog/how-retailers-personalize-the-buy-with-ai
#generativeai #chatbot #nrf2024 -
AI21 Labs
@AI21Labs
Meet us at #nrf24 and see how AI-powered solutions make personalization in retail effortless.
👉 Book a meeting with one of our experts: ai21-labs.webflow.io/national-retail-federation-conference -
AI21 Labs reposted
Brave
@brave
The Brave Technologist Podcast:
Limitations of Artificial Intelligence and Misaligned User Expectations
Gilad Lumbroso (@gullumbroso), Squad Lead for @AI21Labs, discusses the role that natural language will play in the evolution of new AI, and how to apply natural language usability as a key part of product development, on the latest episode of #thebravetechnologist.
Listen to the full podcast, hosted by @LukeMulks, here: kite.link/S8E8 -
AI21 Labs
@AI21Labs
Happy to announce our latest partnership with Invisible Technologies, the industry leader in AI training and support! 🤝
Read more here: www.ai21.com/blog/ai21-invisible-technologies-strategic-p… -
AI21 Labs
@AI21Labs
We’re here in Vegas! ✨ Swing by booth 205 to catch the action—live demos, cool swag, 1:1 sessions, and see how we're shaping the future of LLMs. 🚀
-
AI21 Labs
@AI21Labs
Excited to share our latest partnership with @SnowflakeDB!❄️ Developers can now easily access our Jurassic Models via Snowpark Container Services.
www.ai21.com/blog/ai21-labs-state-of-the-art-language-mod… -
AI21 Labs
@AI21Labs
📣 Reminder that #awsreinvent is just around the corner! Secure your spot and learn how to build optimal generative AI architecture: reinvent.awsevents.com/register/
-
AI21 Labs
@AI21Labs
Welcome Tom Nides, former US Ambassador to Israel, to AI21's Board of Advisors! We look forward to working together and learning from his unique perspective of the financial market.
Glad to have you on board, Tom! 🌟 -
AI21 Labs
@AI21Labs
Our Series C round has not only met but exceeded expectations - with an extension of $53.5 million, bringing the total amount of this round to $208 million! 📈
Special thanks to our latest investors @intelcapital and @ComcastVentures for joining us on this exciting journey. -
AI21 Labs
@AI21Labs
Come meet the team at #awsreinvent in Vegas! 👋
Let's talk #genai - schedule a meeting with us: share-eu1.hsforms.com/14zBeOy_zS7Sbwq1-EKr5QAf8654 -
AI21 Labs
@AI21Labs
"Diversity breeds innovation, and with Dataiku's LLM Mesh, we're stepping into a future of boundless AI possibilities." - Pankaj Dugar, SVP, AI21 Labs.
Jurassic LLMs are now in @dataiku's LLM Mesh ecosystem! Together, we're building a future where #genai is safe and accessible.
AI21 Labs
@AI21Labs
Read more here: www.ai21.com/blog/ai21-labs-announces-partnership-with-da…
-
AI21 Labs
@AI21Labs
Our Jurassic foundation models are widely available in @awscloud Bedrock! We are excited to provide more AWS builders and businesses with our reliable, sophisticated and flexible LLMs.
Read more here >> www.ai21.com/blog/ai21-labs-on-bedrock-now-available-for-… -
AI21 Labs
@AI21Labs
It's official! Our Jurassic LLMs and Task-Specific models are now available on Google Marketplace.
Customers can now easily tap into these resources to build and deploy generative AI-driven applications. 🚀
www.ai21.com/blog/ai21-labs-latest-llms-now-available-in-… -
AI21 Labs
@AI21Labs
Be sure to drop by booth 223 at this year's SaaStr Annual event! Explore the remarkable capabilities of AI21 Studio and discover how you can can easily customize language models to meet your specific needs.
#ai21studio #saastr2023 #nlp -
AI21 Labs reposted
Amnon Shashua
@AmnonShashua
We are at a tipping point in AI – a revolution that would fundamentally change the way we interact with machines. The revolution of natural language understanding, combined with the emergence of “out-of-distribution” reasoning, will be the next great milestone in developing “broad-AI” applications, turning computers into thought partners instead of mere tools. This emergence is what helped us achieve great milestones in such a short time and is powering us to plow ahead full-throttle.
@AI21Labs has always been a pure play in AI, developing and owning foundation models (Jurassic), which serve as a platform play for developers and enterprises while offering our @wordtune suite, used by over 10 million people, which received a major update to its capabilities just days ago.
The current round of funding will allow us to grow and continue to develop the next leap in AI with the capabilities of reasoning across many domains of human content generation. The impact of AI21 Labs' growth plans will be on a global scale and quite soon.
www.ai21.com/blog/series-c-funding/ -
AI21 Labs
@AI21Labs
We're thrilled to announce that we've raised $155 million in series C funding, proudly joining the unicorn club with a valuation of $1.4 billion! 🦄
www.ai21.com/blog/series-c-funding -
AI21 Labs reposted
VentureBeat
@VentureBeat
AI21 Labs debuts Contextual Answers, a plug-and-play AI engine for enterprise data venturebeat.com/ai/ai21-labs-debuts-contextual-answers-a-…
-
AI21 Labs
@AI21Labs
We're excited to introduce Contextual Answers, an API solution where answers are based on organizational knowledge, leaving no room for AI hallucinations. 💭
➡️ www.ai21.com/blog/introducing-contextual-answers -
AI21 Labs
@AI21Labs
"LLMs change the way we can access information and that enables exciting possibilities for retailers." 🛒
🎙 Tune in to find out how LLMs are transforming the eCommerce game: open.spotify.com/episode/0p1EIKdbAYUqZMcYsCTV62
#llm #ecommerce #retail #aibusiness -
AI21 Labs reposted
Clarivate for Academia & Government
@ClarivateAG
Through our partnership with @AI21Labs, our #ai-powered services will empower #researchers and #students with the tools they need to gain insights from the ever-growing volume of #scholarly content: clarivate.com/news/clarivate-announces-partnership-with-a…
-
AI21 Labs
@AI21Labs
What should you keep in mind when transitioning from demo to production? 🤔
Hear from Yuval Belfer, as he delves into the essential considerations for harnessing the power of #generativeai. ⬇️
youtu.be/WgRUFwjRkKQ?t=466 -
AI21 Labs reposted
Daniel Jannai
@DanielJannai
“Human or Not? A Gamified Approach to the Turing Test” by @AI21Labs
We’ve conducted the largest-scale Turing-like test to date, with more than 10M chats and 1.5M participants, and the results are fascinating! 👤🤖
Game: www.ai21.com
Paper: arxiv.org/abs/2305.20010
🧵 -
AI21 Labs
@AI21Labs
It's clear that #genai is becoming a priority for companies everywhere. 🦾
Thank you to @GoogleCloud_IL for hosting this incredible conference, we look forward to many more successful events in the future!
Google Cloud Israel
@GoogleCloud_IL
The #ai vision is becoming a business reality!
Industry leaders from @AI21Labs and Google Cloud had a fascinating discussion on how businesses can utilize Generative AI to enhance and advance their company. 💪
Want to hear more about #generativeai? Stay tuned for more in June! -
AI21 Labs
@AI21Labs
"Humans or Not" has become the largest Turing Test in history, with 1.5 million participants from around the world. 🌎
💡 Uncover the fascinating insights and surprising strategies: www.ai21.com/blog/human-or-not-results -
AI21 Labs
@AI21Labs
Ready to learn from the best? 🎓Join top industry leaders from Google Cloud, and our Co- CEO Prof. Yoav Shoham and developer advocate Yuval Belfer, for a roundtable discussion on generative AI. 🤖
Sign up now to secure your spot: inthecloud.withgoogle.com/integrating-generative-ai-israe… -
AI21 Labs
@AI21Labs
Congratulations to our chairman, Prof. @AmnonShashua, on being awarded the Israel Prize for lifetime achievement! 🏆This is a remarkable accomplishment and truly well-deserved.
#machinelearning #israelprize #lifetimeachievement -
AI21 Labs
@AI21Labs
Honored to be named among the top 3 most promising startups of 2023 by @Calcalist, and proud to be at the forefront of the #genai revolution! Thanks to Calcalist for this incredible recognition.
Read more about the top 50 startups: www.calcalistech.com/ctechnews/article/syjrkom73 -
AI21 Labs
@AI21Labs
Recent evaluations show that our summarize API outperforms OpenAI's models, offering a better user experience and more consistent results. 📈
Learn more about the research and findings 👇www.ai21.com/blog/summarize-api-outperforms-openais-models -
AI21 Labs
@AI21Labs
We’re excited to share our latest innovation with the world and have it launch today on @ProductHunt!
Introducing our latest gaming experiment, “Human or Not?” - the first ever Social Turing Game. 👾 humanornot.ai -
AI21 Labs reposted
Adam Selipsky
@aselipsky
We’re announcing Amazon Bedrock—giving customers the easiest way to build and scale generative AI applications with access to the leading foundation models including @AnthropicAI, @AI21Labs, @StabilityAI and Amazon Titan FMs. Choice rules!
aws.amazon.com/blogs/machine-learning/announcing-new-tool… -
AI21 Labs
@AI21Labs
We're thrilled to announce our partnership with Amazon for their latest service, Bedrock. As our Co-CEO, Ori Goshen, stated, "With Jurassic-2 models and Bedrock, developers can maximize the performance of language tasks while optimizing the cost."
www.ai21.com/blog/announcing-amazon-partnership -
AI21 Labs
@AI21Labs
It was an incredible experience to be a part of the START Global Summit in St. Gallen!
Our Co-CEO, @origoshen, had the opportunity to discuss the future of #generativeai and the AI21 labs journey with the brightest minds and leaders in the industry.
#start23 #ai #innovation -
AI21 Labs
@AI21Labs
📅Mark your calendars for our second AMA session with our Studio experts, Yuval Belfer and Yotam Shinan! They'll be tackling all your burning questions on Jurassic-2, APIs, NLP, Gen AI, and anything else you throw their way!
Join here: discord.com/invite/sKBFxqEx
#ai21studio #ama -
AI21 Labs reposted
SwissCognitive, AI Venture, Advisory & Research
@SwissCognitive
"#ai isn’t here to take over your job, but it is here to optimize your work and make life easier"🦾
We are happy to welcome Ori Goshen, AI21, among our 20 cross-industry AI experts👏
Register & meet the speakers of next week's #virtualconference:
🔗www.linkedin.com/uas/login?session_redirect=https%3A%2F%2… -
AI21 Labs
@AI21Labs
The possibilities of AI in game development are endless. 🎮
Learn how Ubisoft harnessed the power of our AI to revolutionize their game creation process and achieve more in less time: www.ai21.com/blog/ubisoft-case-study
#gamedevelopment #ai #llm -
AI21 Labs
@AI21Labs
Don't miss the chance to join our own Leeor Moses-Voronov next week for a unique panel discussion on #genai, exploring its possibilities and market concerns from different perspectives.
Register here to join in-person or virtually: docs.google.com/forms/d/e/1FAIpQLSdqwhraP9t-FOs-TFU_jO3Qz… -
AI21 Labs
@AI21Labs
Exciting news! 📣 We've launched Jurassic-2 and our task-specific APIs! Both are major game-changers, with Jurassic-2's new and improved capabilities, and our APIs' plug-and-play reading & writing functions that outperform competitors.
Read more: www.ai21.com/blog/introducing-j2 -
AI21 Labs
@AI21Labs
Catch our Co-CEO @origoshen in the closing panel at the AWS Startup Showcase on March 9th! They'll be discussing the hottest trend in tech - the future of Generative AI 🔥 Join this virtual event here: aws.events.cube365.net/startup/aws-startup-showcase-s3-e1…
#generativeai #techtrends #leadershipadvice -
AI21 Labs reposted
Amnon Shashua
@AmnonShashua
1/4
With #ai, authors such as Eshkol Nevo will be able to write dozens of books every year. Yesterday at Yediot’s ‘People of The Year’ conference, we had an interesting talk about the role AI will play in the realm of content creation -
AI21 Labs
@AI21Labs
Curious to learn more about our models? Join our upcoming AMA session this Wednesday, hosted by our NLP experts and Developer Advocates Yotam Shinan and Yuval Belfer.
Head over to our Discord, "AMA #1," and get ready to ask us your burning questions 🔥 discord.com/invite/7YqhxMVS -
AI21 Labs
@AI21Labs
We are beyond excited to have won Startup of the Year at the #geektimeawards! A huge thank you to our amazing team and to @geektime for this honor and recognition. This is only the beginning for us! 🏆
-
AI21 Labs
@AI21Labs
Excited to share our latest case study on Verb, a new AI writing tool powered by AI21 Studio!
With the use of our custom models, Verb goes beyond mere text generation and becomes a true partner to authors everywhere.
www.ai21.com/blog/verb-ai-case-study -
AI21 Labs
@AI21Labs
Interview with our Co-Chief Scientist, @YoavLevine: "When someone creates content, they want to be comfortable to put their name on it. And in certain use cases, the ability to connect the content to sources really facilitates this” bdtechtalks.com/2023/01/30/ai21labs-llms-ralm/ via @bdtechtalks
-
AI21 Labs reposted
Ori Ram
@ori__ram
#nlproc
New paper!
"In-Context Retrieval-Augmented Language Models"
From @AI21Labs 🚀
You can get very large LM gains by concatenating retrieved documents to the LM input *without any further training*
Paper: uploads-ssl.webflow.com/60fd4503684b466578c0d307/63d92258…
Code (tmrw): github.com/AI21Labs/in-context-ralm
🧵 -
AI21 Labs
@AI21Labs
We are proud to announce the release of our latest scientific paper about improving grounded text generation and fact attribution - for any off-the-shelf LMs!
Read more details about this exciting development and its implications for the future of #ai: www.ai21.com/blog/grounding-language-models-in-context -
AI21 Labs
@AI21Labs
"Generative anything is the idea that you put in a little and get a lot. It's exciting as it is wondrous"
Check out this recent CTech interview with our Co-Founder & Co-CEO, Prof. Yoav Shoham: www.calcalistech.com/ctechnews/article/bkgetacys -
AI21 Labs reposted
Amnon Shashua
@AmnonShashua
. @wordtune is going mobile.. once productizing generative AI as a writing tool will be completed, writing will never be the same… more to come from @AI21Labs
www.jpost.com/business-and-innovation/tech-and-start-ups/… -
AI21 Labs
@AI21Labs
Our friends at Verb have launched their AI-enhanced writing apps for authors. It's been a pleasure working with their wonderful team! 🥳
Verb
@verbforwriters
Today, we’re thrilled to launch Verb, an AI-enhanced writing app for fiction writers! ✍️🤖
Verb helps you write, brainstorm, plan, and get feedback on your work.
Sign up for free at verb.ai!
Read on for more on our thinking 👇🧵 -
AI21 Labs
@AI21Labs
Our Co-Founder & Co-CEO, @yshoham will participate in @EntreeCap's panel discussion on #generativeai. You are welcome to join!
-
AI21 Labs
@AI21Labs
Generative AI continues to rapidly develop, creating a new reality. But we're only just beginning and there are few limitations. Read more about it in @Tech12news, recent Op-ed written by our own @DanPadnos (HEB)
www.tech12.co.il/index-commentary/Article-4b0361cf178e481… -
AI21 Labs
@AI21Labs
Have you heard the news? we're now offering Jurassic-1 Grande on Amazon SageMaker! Check out the blog post and sign up for the preview:
aws.amazon.com/blogs/machine-learning/ai21-jurassic-1-fou…
#generativeai #aws #largelanguagemodels #nlp #ai
Dan Padnos
@DanPadnos
Great #reinvent2022. Really enjoyed co-speaking with @GalOshri and Emily Webber in our session on distributed training and LLMs yesterday. Monday we announced that @AI21Labs Jurassic-1 is now available on @awscloud SageMaker. And today we’re already hearing from eager customers!
-
AI21 Labs
@AI21Labs
Our Co-founder & Co-CEO @origoshen kicking off @awscloud #reinvent2022 with an exciting announcement: you can now deploy our LLM Jurassic-1 on your private SageMaker environment!
-
AI21 Labs reposted
Eran Yacobovitch
@eranyac
לא הייתי כל כך פעיל כאן לאחרונה, עבודה חדשה וכאלה. אבל חשבתי שעם כל הבאזז סביב ג׳נרטיב AI (וכל עוד טוויטר עדיין קיימת) זה זמן מעולה לתת מבוא קצר ולשתף איך הדברים נראים ״מבפנים״.
אז יאללה שרשור #generativeai 🧵⛓️ >> -
AI21 Labs reposted
The Pulse of AI
@ThePulseOfAI
Latest @thepulseofai #podcast with @ai21labs co-founder Dr. Yoav Shoham www.thepulseofai.com/post/conversation-with-dr-yoav-shoha…
-
AI21 Labs
@AI21Labs
Come visit us at the upcoming IL Cloud Summit on November 9th!
We'll be discussing a topic that has been gaining plenty of traction - large language models.
Be sure to stop by our booth as well to experience the magic of our AI fortune teller 🔮✨ -
AI21 Labs
@AI21Labs
Will AI eventually replace writers in the future?
We're not fortune tellers but our crystal ball says...no. 🔮
We don't believe it will, but here's how it may change the industry in some truly revolutionary ways:
www.wordtune.com/blog/future-ai-writing -
AI21 Labs
@AI21Labs
Our co-founder Yoav Shoham was recently featured on 20 Minute Leaders, a daily talk show that hosts thought leaders from around the world 🌎
Check out the full podcast on Spotify: open.spotify.com/episode/2Wmu0y2besQLZxVPxkvFNm -
AI21 Labs
@AI21Labs
Big thanks to our friends over @GoogleCloud_IL for sending us this sweet surprise after our latest round of funding!
Here's to a great partnership and many more celebratory treats to come! 🎂 -
AI21 Labs
@AI21Labs
AI21 Labs has raised $64 million in series B funding, bringing our valuation to $664 million!
Thanks to Ahren Innovation Capital for leading the round, we look forward to growing our incredible #ai21team and investing more in R&D.
Read more here: techcrunch.com/2022/07/12/openai-rival-ai21-labs-raises-6… -
AI21 Labs reposted
Dan Padnos
@DanPadnos
סיכום מעולה של הדברים המדהימים שקורים פה ב-@AI21Labs. בואו גם!
calcalist360.webflow.io/articles/ai21lab -
AI21 Labs
@AI21Labs
You're Welcome. ask-rbg.ai
Elisa Gabbert
@egabbert
If science was worth anything it would be building a Ruth Bader Ginsburg AI
-
AI21 Labs
@AI21Labs
“There are not many places where the general public can go and play with real AI, but now you can.”
This Washington Post article explores our most recent AI experiment: www.washingtonpost.com/technology/2022/06/14/ruth-bader-g… -
AI21 Labs
@AI21Labs
What would RBG (probably) say about… anything? Check out our latest AI experiment at ask-rbg.ai and let her judge! #askrbg
-
AI21 Labs
@AI21Labs
Today! Twitter Spaces discussion on best practices for deploying LLMs, led by @percyliang with our very own
@Udi73613335 and other industry partners. See you there!
Percy Liang
@percyliang
Join us tomorrow (Wed) at 12pm PT to discuss the recent statement from @CohereAI @OpenAI @AI21Labs on best practices for deploying LLMs with @aidangomezzz @Miles_Brundage @Udi73613335. Please reply to this Tweet with questions! x.com/i/spaces/1vAxRkqBVmXKl
-
AI21 Labs
@AI21Labs
Best practices when deploying large language models. Joint statement with @OpenAI and @cohereAI www.ai21.com/blog/best-practices-for-deploying-language-m…
-
AI21 Labs
@AI21Labs
To be perfectly Queer, EVERY month is #pride_month. Enriching Expression is kinda our thing at @Wordtune. So, let’s celebrate #lgbtq+ rights & expression.
Flaunt your truth, and celebrate all that is fabulous 🌈👭🌈 -
AI21 Labs
@AI21Labs
We’ve been named one of the top 100 most innovative AI startups of 2022 by @CBinsights !
We are proud that our language models take part in pioneering this momentous period in AI history, and we look forward to shaping the future of NLP as an intelligent thought partner. -
AI21 Labs
@AI21Labs
Our very own @GalShachaf will be presenting at #awssummit in Tel Aviv next week to speak about what we're developing here at AI21 Labs and how we use Amazon SageMaker to train large language models. You can register for the summit in this link: go.aws/3sqPOVQ
-
AI21 Labs
@AI21Labs
1/n Introducing J1-Grande, the newest member of the Jurassic-1 family. With 17B parameters, Grande offers Jumbo-like quality (10x larger) at an affordable price (⅓ of the cost).
AI21 Labs
@AI21Labs
2/n We built a poet and didn’t even know it. We requested a witty poem about the relationship between feline and canine. Here is what the J-1 master sonneteers composed:
AI21 Labs
@AI21Labs
3/n Ready to meet the new kid on the block? Check out the blog post: www.ai21.com/blog/introducing-j1-grande
-
AI21 Labs reposted
Dan Padnos
@DanPadnos
The team at @AI21Labs is churning out some cool science. We now see that large pre-trained language models (LLMs) have even more potential than anyone imagined. My favorite result in the paper: Outperform T0++ on multi-task zero-shot WITHOUT fine-tuning an LLM! More here 👇
Yoav Levine
@YoavLevine
1/n New paper: Standing on the Shoulders of Giant Frozen LMs ❄️❄️ -- you can get much more out of a single frozen language model! 🔥🔥 Check it out in our new preprint 🤓 storage.cloud.google.com/ai21-public-data/publications/St…
-
AI21 Labs reposted
Yoav Levine
@YoavLevine
1/n New paper: Standing on the Shoulders of Giant Frozen LMs ❄️❄️ -- you can get much more out of a single frozen language model! 🔥🔥 Check it out in our new preprint 🤓 storage.cloud.google.com/ai21-public-data/publications/St…
-
AI21 Labs
@AI21Labs
A language model that can talk to APIs? 😮 Say hello to Jurassic-X! 🦕🦕🦕 Read more on our blog, play with the demo, and sign-up for early access
www.ai21.com/blog/jurassic-x-crossing-the-neuro-symbolic-…Post not available here, view on X
AI21 Labs
@AI21Labs
Anddd knows to how to route questions to an external database 🚢🧊
AI21 Labs
@AI21Labs
Big language models just got smarter 🧠 Contact us for early access here:
dghsiym3kjj.typeform.com/to/rnIzTBQ0 -
AI21 Labs
@AI21Labs
1/n Hot off the press – AI21 Studio graduated from Beta!! Seamlessly train your own custom model by the touch of a button using our enhanced user interface.
AI21 Labs
@AI21Labs
2/n Enjoy all models with no strings attached: pay-as-you-go to taste-test them all.
AI21 Labs
@AI21Labs
3/n Running out of no-show excuses? Let AI generate a convincing and rather witty ticket to freedom: youtu.be/kc6qRZZFDKQ
AI21 Labs
@AI21Labs
4/n We’ve seen it all. Hit us with a seemingly-complex Use Case, and sit back as Studio trains a custom model flawlessly meeting your every need. youtu.be/DjhrixcO6cM
AI21 Labs
@AI21Labs
5/n And for my fellow inquisitive humans, check out the Studio How-To here: studio.ai21.com/docs/getting-started/. Alright enough from us– go play on our Playground. Fair warning: its genius is addicting.
-
AI21 Labs reposted
Google Cloud Israel
@GoogleCloud_IL
“Women’s day is a great opportunity to cherish all the amazing women in our lives that supported us along the way” @gorvitron, director of partnership & special projects at AI21 Labs. Find out more about Olivia’s thoughts on #internationalwomensday ↓ goo.gle/3Djnati
-
AI21 Labs
@AI21Labs
Yoav Shoham (@yshoham), our Co-Founder & CEO in an interview with Young Sohn, discussing AI21 Labs – which Young’s Walden Catalyst is an investor in – and its solutions, the history and future of artificial intelligence (AI).
youtu.be/MxHjzExC3nQ -
AI21 Labs
@AI21Labs
Our co-founder & co-CEO @yshoham was chosen as one of Israel's most influential leaders in deep tech! Congrats Yoav and lucky us :)
www.tech12.co.il/index-technology_first/Article-13fdc547f… -
AI21 Labs
@AI21Labs
Spotted! Our awesome Co-CEO @origoshen on a billboard (!!!) in Tel Aviv. Thanks @Google - we ❤️ you too.
-
AI21 Labs reposted
Dan Padnos
@DanPadnos
Check out @Alber_RomGar 's piece on @AI21Labs and our 178B-parameter language model, Jurassic-1 Jumbo.
Great time chatting with you, Alberto!
Towards Data Science
@TDataScience
In his latest exploration of AI technologies, @Alber_RomGar tinkers with J1-Jumbo, AI21's 178-billion-parameter language model. towardsdatascience.com/how-ai21-labs-is-reimagining-the-w…
-
AI21 Labs reposted
AI Dungeon
@aidungeon
We’re rolling out significant updates to language models in #aidungeon (starting with 10% of users, gradually increasing over the week). More details in our blog: latitude.io/blog/update-language-models
-
AI21 Labs reposted
Wordtune
@wordtune
OMG @googlechrome just picked Wordtune as one of their favorite extensions of 2021! We want to share this award with our awesome users - couldn't have done it without ya! #chromewebstorefavorites
We're in great company! @Rakuten @KamiApp @loom & more: chrome.google.com/webstore/category/collection/2021_favor… -
AI21 Labs
@AI21Labs
We <3 @GoogleCloud_IL!
Google Cloud Israel
@GoogleCloud_IL
6/ It’s our customers who are the real stars🌟
They take our solutions and create amazing things. We have the privilege to hear from @origoshen @Wordtune by @AI21Labs ,Yaron Inger @Lightricks, and Tzury Bar Yochay @Reblaze about how Google Cloud is helping them innovate faster. -
AI21 Labs reposted
Wordtune
@wordtune
Welcome to the family, Wordtune Read. A year ago we launched Wordtune to empower people across backgrounds, nationalities and industries to write with more confidence and ease. Now with Wordtune Read we hope to do the same for reading!
Try it out! wordtune.com/read -
AI21 Labs reposted
Walden Catalyst
@walden_catalyst
Our founding managing partner Young Sohn talked to @Geektime about his views on the Israeli market and @walden_catalyst latest investment in @AI21Labs, founded by @yshoham, @origoshen, and @AmnonShashua.
geektime.com/walden-catalyst-and-former-samsung-president…
#israel #deeptech #ai -
AI21 Labs reposted
CTech
@Calcalistech
.@AI21Labs launches AI-based tool that summarizes a document within seconds.
The Israeli startup, which recently raised $20M at a $400M valuation, believes it can enable users to read long and complex text more efficiently and on average 75% faster.
www.calcalistech.com/ctech/articles/0,7340,L-3922633,00.html -
AI21 Labs reposted
Last Week in AI
@Last_Week_in_AI
GPT-3 is amazing, but sadly @OpenAI opted to not release it for others to build upon.
1.5 years later, things have changed: more and more models akin to GPT-3 are getting developed, and in some cases released.
Read all about that in our latest:
👇
lastweekin.ai/p/gpt-3-is-no-longer-the-only-game -
AI21 Labs reposted
AI21 Labs
@AI21Labs
Our VP of Platform @DanPadnos is speaking at @JohnSnowLabs NLP Summit this Thurs, Oct 7th at 2p EST. Come learn more about AI21 Studio & Jurassic-1!
www.nlpsummit.org/zero-to-production-bootstrapping-text-b… -
AI21 Labs
@AI21Labs
Our VP of Platform @DanPadnos is speaking at @JohnSnowLabs NLP Summit this Thurs, Oct 7th at 2p EST. Come learn more about AI21 Studio & Jurassic-1!
www.nlpsummit.org/zero-to-production-bootstrapping-text-b… -
AI21 Labs
@AI21Labs
We're in open beta! Try out Jurassic-1 now: studio.ai21.com
venturebeat.com/2021/09/11/stuck-in-gpt-3s-waitlist-try-o… -
AI21 Labs reposted
Amnon Shashua
@AmnonShashua
A transformative moment in the evolution of #ai – we released Jurassic-1 – the largest #nlp model in the world – to the community. Now any developer can build language applications with just a few examples. @AI21Labs @yshoham @origoshen www.zdnet.com/article/watch-out-gpt-3-here-comes-ai21s-ju…
-
AI21 Labs
@AI21Labs
We love this Jurassic response. Share all of your fun/ quirky/enlightening Jurassic interactions with us! #chatswithjurassic
Noam Siegelman
@NoamSiegelman
I saw the new jumbo model by @AI21Labs and wondered whether it will be useful for me as a cognitive scientist working on human language. Because it is so smart I decided to ask it directly - the answer was suprisingly modest 😂
-
AI21 Labs reposted
CogX Festival
@CogX_Festival
Have @OpenAI met their match in #nlp service with rising star, @AI21Labs? @VentureBeat thinks so.
venturebeat.com/2021/08/11/ai21-labs-trains-a-massive-lan…
Unfamiliar with NLP? Why not check out our #cogx2021 talk with experts @Franki_hackett & Ioannis Efstathiou of @EngineB_ 👇
www.youtube.com/watch?v=punecIvSYSg&hss_channel=tw-478464… -
AI21 Labs reposted
Elman Mansimov
@elmanmansimov
Played a bit with @AI21Labs language models to build a conversational semantic parser on fast-food domain with 50 prompts.
It worked well...
And surprisingly generalized to ordering things beyond food such as t-shirt and TV.
Wish I didn't exceed 175B model quota that fast 🙃 -
AI21 Labs
@AI21Labs
We're shifting the paradigm of how to scale text-based AI applications affordably & efficiently. Read about our custom models approach and learn how to take your app from prototype to production.
www.ai21.com/blog/zero-to-production-in-ai21-studio -
AI21 Labs reposted
Fei-Fei Li
@drfeifei
Congrats to @yshoham and colleagues!
AI21 Labs
@AI21Labs
Today we’re launching AI21 Studio, a platform where you can instantly access our state-of-the-art language models to build your own applications - including a 178B parameters model, Jurassic-1 Jumbo. We can’t wait to see what you create!
www.ai21.com/blog/announcing-ai21-studio-and-jurassic-1 -
AI21 Labs reposted
Philip Vollet
@philipvollet
Competing GPT-3: Bigger is always better? Nope! But say hello to AI21 Labs’ Jurassic-1 a 178B-parameter Language Model and get instant access and build sophisticated text-based AI applications at scale
@AI21Labs
Web www.ai21.com/blog/announcing-ai21-studio-and-jurassic-1
Paper as PDF uploads-ssl.webflow.com/60fd4503684b466578c0d307/61138924… -
AI21 Labs reposted
Erik Brynjolfsson
@erikbryn
.@AI21Labs is launching a language model even bigger than GPT-3. These are exciting times!
AI21 Labs
@AI21Labs
Today we’re launching AI21 Studio, a platform where you can instantly access our state-of-the-art language models to build your own applications - including a 178B parameters model, Jurassic-1 Jumbo. We can’t wait to see what you create!
www.ai21.com/blog/announcing-ai21-studio-and-jurassic-1 -
AI21 Labs reposted
Eze Vidra
@ediggs
Israeli startup @AI21Labs launched a free GPT-3 equivalent with no waitlist and publicly available
AI21 Studio provides access to Jurassic-1, our suite of language models that includes Jurassic-1 Jumbo, the largest model ever made available to developers.
www.ai21.com/studio -
AI21 Labs
@AI21Labs
We used our Jurassic-1 language models to generate rap lyrics about @BarackObama. We're basically the next @Lin_Manuel, right? www.youtube.com/watch?v=vO3xvX0ry9M
Discover tons of other ways to use AI21 Studio from the frivolous to the substantial: www.ai21.com/blog/ai21-studio-use-cases -
AI21 Labs reposted
Paras Chopra
@paraschopra
Just tried this model.
It's *really* good. Questions are mine, answers are what the model generates.
Great job @AI21Labs! -
AI21 Labs reposted
Spiros Margaris
@SpirosMargaris
Watch out, #gpt3,
here comes #ai21's 'Jurassic' #languagemodel
www.zdnet.com/article/watch-out-gpt-3-here-comes-ai21s-ju… #fintech #ai #artificialintelligence #machinelearning #deeplearning @TiernanRayTech @ZDNet -
AI21 Labs reposted
Kevin Leyton-Brown
@k_leyton_brown
@AI21Labs announces a GPT-3 successor that anyone can use! The model is bigger (178B params) and finally, publicly available. If you aren't familiar with breakthroughs in language modeling over the past couple years, you'll be AMAZED. AI is getting closer. studio.ai21.com
-
AI21 Labs reposted
ZDNET
@ZDNET
Watch out, GPT-3, here comes AI21's 'Jurassic' language model www.zdnet.com/article/watch-out-gpt-3-here-comes-ai21s-ju…
-
AI21 Labs reposted
Mark Riedl
@mark_riedl
.@AI21Labs announces an open beta of a 178B parameter language model www.ai21.com/blog/announcing-ai21-studio-and-jurassic-1
-
AI21 Labs reposted
/MachineLearning
@slashML
AI21 releases a 178B parameter model in par with GPT-3 for free usage! www.reddit.com/r/MachineLearning/comments/p2l54l/n_ai21_r…
-
AI21 Labs reposted
Bob E. Hayes
@bobehayes
Two new #ai models can read, write — and turn words into computer code
www.axios.com/openai-ai21-labs-natural-language-process-a…
#nlp #artificialintelligence @AI21Labs @OpenAI -
AI21 Labs reposted
Fedor Kovalev
@KesselmanF
Wanted to try GPT-3 but was stuck on the waitlist?
Today @AI21Labs announced Studio, NLP-as-a-service platform for SMBs, solopreneurs, and builders.
The interface is similar to OpenAI and it was able to generate decent results for most templates Magic Educator already has. -
AI21 Labs
@AI21Labs
Check out our white paper for a comprehensive technical description and thorough evaluation of our Jurassic-1 (J1) language models: uploads-ssl.webflow.com/60fd4503684b466578c0d307/61138924…
-
AI21 Labs reposted
ZDNET
@ZDNET
Watch out, GPT-3, here comes AI21's 'Jurassic' language model zd.net/3AvbjGb by @TiernanRayTech
-
AI21 Labs reposted
VentureBeat
@VentureBeat
AI21 Labs trains a massive language model to rival OpenAI's GPT-3 venturebeat-com.go-vip.co?p=2709686 by @Kyle_L_Wiggers
-
AI21 Labs
@AI21Labs
"Developers will be able to sign up for free to try the service, & can then apply for access to custom models for commercial aims "to use our technology to quickly build any text-based applications they can dream up," says @DanPadnos." @bryanrwalsh @axios
www.axios.com/openai-ai21-labs-natural-language-process-a… -
AI21 Labs
@AI21Labs
Today we’re launching AI21 Studio, a platform where you can instantly access our state-of-the-art language models to build your own applications - including a 178B parameters model, Jurassic-1 Jumbo. We can’t wait to see what you create!
www.ai21.com/blog/announcing-ai21-studio-and-jurassic-1 -
AI21 Labs
@AI21Labs
Check out our presentation of “PMI-Masking: Improving LM training by masking correlated spans” at #iclr2021!
Oral session: Mon 05/03 2PM EST
iclr.cc/virtual/2021/session/4323
Poster session: Thu 05/06 12PM EST
iclr.cc/virtual/2021/papers.html?filter=titles&search=PMI…
See our blogpost on PMI-Masking
www.ai21.com/pmi-masking -
AI21 Labs reposted
NoCamels - Israeli Tech and Innovation News
@NoCamels
. @AI21Labs, a company established by AI luminaries and veterans of the IDF elite 8200 tech unit, has launched the first AI writing companion that understands context and meaning. nocamels.com/2021/01/ai21-labs-wordtune-writing-tech-ai/
-
AI21 Labs
@AI21Labs
Today we announced the launch of our first product @wordtune, an #ai based tool that understands both writing & meaning. Check out our new writing assistant on @googlechrome, and read more on the announcement from @jeremyakahn in @FortuneMagazine: fortune.com/2020/10/27/editing-tool-wordtune-a-i/
-
AI21 Labs
@AI21Labs
Happy to share this wine story published on Jancis Robinson, featuring HAIMKE and created by Prof. Yoav Shoham
www.jancisrobinson.com/articles/are-wine-writers-redundant -
AI21 Labs
@AI21Labs
Check out @yshoham's keynote speech from #gmic
www.ai21.com/nlp-gmiclivepro -
AI21 Labs
@AI21Labs
Our own @yshoham will be keynoting at #gmic on Sep 25.
thegmic.com -
AI21 Labs
@AI21Labs
We’re excited to share Auxiliary Tuning - a simple and efficient method for reducing training costs by adapting a pre-trained LM to a novel task, e.g. conditional text generation. www.ai21.com/auxiliary-tuning
#nlproc #deeplearning -
AI21 Labs reposted
Kevin Leyton-Brown
@k_leyton_brown
Amazing new AI writing demo from our team at @AI21Labs: you write bullet points; it generates text. Definitely not ready for prime time, but points the way to word processor 2.0: tell a computer what you want to say instead of writing directly. Try it out: www.ai21.com/haimke
-
AI21 Labs
@AI21Labs
@timshiel we really liked how you used #haim in your radio show! We released a new text generator today called HAIMKE. Give it a try! we'd love to see what you create www.ai21.com/haimke
AI21 Labs
@AI21Labs
@timshiel if you enjoyed #haimke, we'd appreciate if you could vote for us on Product Hunt: www.producthunt.com/posts/write-with-haimke
-
AI21 Labs
@AI21Labs
@jonathanfly we remember you enjoyed #haim, our interpolating LM. We just released HAIMKE, a new language model that turns bullet points to a full document - tons of fun! Give it a spin, we're excited to see what you come up with: www.ai21.com/haimke
AI21 Labs
@AI21Labs
@jonathanfly if you enjoyed using #haimke, we'd appreciate if you could vote for us on Product Hunt: www.producthunt.com/posts/write-with-haimke
-
AI21 Labs reposted
Ido Klein עדו קליין
@idoklein1
Check out what HAIMKE AI wrote based on my bullet points! haimke-static.ai21.com/html/40807ccac0bb467e83a38d3b0749c… #madewithhaimke via @AI21Labs
-
AI21 Labs
@AI21Labs
We just released HAIMKE on Product Hunt! A neural language model that turns your bullet points into full-fledged text. Give it a try and share some ❤️ www.producthunt.com/posts/write-with-haimke #madewithhaimke #haimke #ai #nlproc @ai21labs
-
AI21 Labs reposted
AIM
@Analyticsindiam
The researchers estimate that, at list-price, training the 11B parameter variant of T5 costs well above $1.3 million for a single run!
analyticsindiamag.com/deep-tech/deep-learning-costs-cloud…
@yshoham @AI21Labs
#deeplearning -
AI21 Labs
@AI21Labs
We just posted a brief overview about the cost of training #nlp language models: arxiv.org/abs/2004.08900
@Or_Sharir @pelegbarak @yshoham -
AI21 Labs reposted
Vasileios Lioutas
@billiout
According to the following study (arxiv.org/abs/2004.08900) from @AI21Labs, training a single BIG NLP model can cost about $10k. That's unacceptable! Both for the environmental burden as well as for the independent researchers who don't have access to these resources.[1/2]
#nlproc
Vasileios Lioutas
@billiout
Excited about our new paper on sequence learning. In the NLP community, attention is considered a necessity for achieving SoTA results but we propose a fast novel attention-free method with comparative performance arxiv.org/abs/2002.03184
#nlproc #deeplearning #machinelearning -
AI21 Labs reposted
Denny Britz
@dennybritz
Just how much does it cost to fully train an NLP model?
Report: arxiv.org/abs/2004.08900
According to this, using non-preemptible AWS/GCP instances and no discounts, you’re looking at $50K to $10M with hyperparam optimization, depending on model size. -
AI21 Labs reposted
Verdict
@VerdictUK
Autonomous vehicles received the largest percentage of private artificial intelligence investment for 2018 and 2019. @StanfordHAI @McKinsey @AI21Labs @Genpact @OpenAI @GoogleAI www.verdict.co.uk/private-ai-investment-autonomous-vehicles/
-
AI21 Labs
@AI21Labs
AI21 Labs is proud to be a sponsoring partner of @IndexingAI. Check it out! Go @yshoham
AI Index
@indexingai
One of the most comprehensive studies about AI is now available. #aiindex2019 tracks and visualizes AI-related information to help ground conversations in data. Read more: hai.stanford.edu/news/introducing-ai-index-2019-report @StanfordHAI
-
AI21 Labs reposted
Robots on Typewriters
@RobotTypewriter
Check out the HAT3000 project on @JanelleCShane’s blog aiweirdness.com and don’t forget to play around with HAIM by @AI21Labs!
-
AI21 Labs reposted
Jose Camacho Collados
@CamachoCollados
It's almost as if there's a pattern here...🤔😂 #bert #wic #nlproc
-
AI21 Labs
@AI21Labs
arxiv.org/abs/1908.05646 SenseBERT: Pre-training with a novel self-supervised word sense prediction task improves #bert's lexical semantics abilities, achieving a state-of-the-art score on the notorious Word-in-Context task. @tpilehvar #superglue #wordnet #nlproc #deeplearning















































































































































































































































































