AI21 Labs builds Foundation Models and AI Systems for the enterprise that accelerate the use of GenAI in production. Meet AI21 Maestro ai21.com/
- Tweets 655
- Following 161
- Followers 11,071
- Likes 687
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 ai21.com/jamba
🔨Build on @huggingface
35
240
1,095
332,801
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 ↓
🔗 ai21.com/blog/rise-of-hybrid…
12
93
467
49,974
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
105
96
415
165,058
📄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 👇1/7
6
77
292
23,738
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!
ai21.com/blog/announcing-ai2…
18
79
271
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: ai21.com/blog/announcing-jam…
(1/4)
10
55
272
42,038
📄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 👇1/6
3
59
250
23,713
🚀 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: ai21.com/blog/structured-rag…
👇Dive into the paper: arxiv.org/abs/2511.08505v1
#RAG #LLMs #StructuredRAG #AI21Labs
6
33
233
35,809
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.
5
28
202
409,999
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: ai21.com/jamba/
6
64
173
31,256
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
2
24
162
21,679
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
108
6
62
11,551
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
ai21.com/blog/jurassic-x-cro…
3
29
113
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! 🦄
ai21.com/blog/series-c-fundi…
8
20
98
20,938
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! ai21.com/maestro
#AI21Maestro #AccuracyAtScale #AIOrchestrator #AIPlanning
1
25
96
170,883
Come join us at #SFTechWeek! Sign up now: lu.ma/nho5h281
🎉 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👇
66
2
32
19,193
Jamba Mini 1.6 outpacing Gemini 2.0 Flash, GPT-4o mini, and Mistral Small 3 on output speed.
artificialanalysis.ai/models…
This is what fast looks like! #BuildOpen #AI21Jamba
2
2
65
143,734
🥂Meet Jamba
ai21.com/jamba
🔨Build with Jamba
huggingface.co/ai21labs/Jamb…
📝Read more
ai21.com/blog/announcing-jam…
#Jamba #Mamba #Transformer #AIarchitecture #SSM
1
8
66
7,629
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: ai21.com/blog/introducing-j2
2
11
49
6,902
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.
1
6
48
10,854
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…
2
6
51
4,004
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
4
9
47
6,808
Best practices when deploying large language models. Joint statement with @OpenAI and @cohereAI ai21.com/blog/best-practices…
1
11
48
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. >>
21
4
26
10,333
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
7
12
45
20,203
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/op…
3
4
42
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 >> ai21.com/blog/ai21-labs-on-b…
1
8
35
21,400
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
2
6
35
5,675
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). >>
34
1
13
7,322
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: ai21.com/blog/grounding-lang…
1
11
36
3,759
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: deeplearning.ai/short-course…
17
3
24
2,743
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.
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.
1
32
2,513
Find our new model on @huggingface.
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: ai21.com/blog/announcing-jam…
Model: huggingface.co/ai21labs/Jamb…
3
25
5,989
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).
2
11
30
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.
1
7
28
2,794
And 140K context on a single GPU. @ArtificialAnlys We look forward to seeing those benchmarks!
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 👀
1
24
1,635
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
1
1
28
2,373
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! 🏆
1
4
25
5,035
Jamba-1.5-Mini's architecture is the same as Jamba. We scaled up the architecture for our Large model. Our technical paper will be posted on arXiv.org in the next few hours.
1
1
16
2,948
“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 👇 ai21.com/events/ai-and-the-f…
3
8
24
60,220
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_groun…
2
26
4,805
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."
ai21.com/blog/announcing-ama…
5
24
4,131
This Jamba overview is a great way to quickly understand the novel features Jamba brings to the dev community. Thanks @AiFlux!
3
23
2,862
#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: eu1.hubs.ly/H09Ln9j0
1
1
22
1,581
5/5 Available today for download & local inference on @huggingface, @kaggle, @lmstudio, and llama.cpp. huggingface.co/ai21labs/AI21…
1
21
938
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: calcalistech.com/ctechnews/a…
1
2
20
20,980
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
21
1,170
We ran a head-to-head #latency test, with the same hardware and same prompts.
Want to guess who won? 🤔
2
20
1,818
High throughput itself is never enough; it's all about optimizing the speed-cost-quality triangle. Thanks to Jamba’s architecture, we offer high speed at a very competitive price (in the image below, QII is where you want to be). >>
21
6
548
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 ❤️ producthunt.com/posts/write-… #madewithHAIMKE #HAIMKE #ai #nlproc @ai21labs
9
17
What would RBG (probably) say about… anything? Check out our latest AI experiment at ask-rbg.ai/ and let her judge! #AskRBG
1
6
19
(2/4) Jamba-Instruct outperforms or rivals other instruction-tuned competitors across common performance benchmarks.
1
19
4,968
We're excited to introduce Contextual Answers, an API solution where answers are based on organizational knowledge, leaving no room for AI hallucinations. 💭
➡️ ai21.com/blog/introducing-co…
3
17
1,190
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! 🎂
2
18
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?
2
17
5,575
(4/4) With a 256K context window, Jamba-Instruct boasts the largest context window in its size class.
3
17
1,833
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 👇ai21.com/blog/summarize-api-…
1
3
17
1,475
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.
4
17
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/produc…
#VertexAI #GoogleCloud #GoogleCloudPartner
3
3
16
5,083
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 --> eu1.hubs.ly/H0bJRt70
1
6
16
4,151
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: ai21.com/blog/karpathys-leas… 🧵 1/6
2
5
17
2,278
It started with the original Mamba paper (Dec 2023) from @_albertgu & @tri_dao:
→ Linear-time inference
→ Content-aware computation
→ Attention-free modeling
That single paper cracked open a whole new path for scalable LLMs.
📄 arxiv.org/abs/2312.00752
1
16
1,290
"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: ai21.com/blog/human-or-not-r…
2
2
14
2,030
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.
1
3
16
(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.
2
1
16
2,174
To support efficient serving of Jamba-1.5-Large, we developed a novel quantization technique - ExpertsInt8. We quantize the MoE and MLP weights to INT8 in order to store them, and dequantize them back to BF16 before the actual computation. This technique is both very fast and without a loss in quality. It also reduces the memory footprint while still enables enjoying the benefits of fast BF16 kernels. 3/7
2
1
15
600
Want to test it yourself? Both models are available at 🤗 Check them out: huggingface.co/collections/a…
9
1
10
5,900
🧠 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…
1
2
16
1,256
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: ai21.com/blog/rise-of-hybrid…
1
12
1,333
📄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
3
13
1,380
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: bit.ly/37HGRxr
4
14
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!
1
9
1,688
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]
1
15
2,097
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
15
1,568
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
2
12
2,596
The hybrid Jamba architecture enables Jamba-1.5 models to reach excellent throughput and latency, especially at long contexts. With the same hardware, Jamba-1.5 models are the fastest across the board (in the image: 2xA100 80GB GPUs for Mini, 8xA100 80GB GPUs for Large). 2/7
1
1
12
696
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/a… via @bdtechtalks
6
14
1,461
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/a…. We can’t wait to see what the community will build on this technology 🤗 7/7
1
13
2,302
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
3
13
1,540
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!
1
14
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
7
14
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
🔗 eu1.hubs.ly/H0bTry50
13
5,275
"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: calcalistech.com/ctechnews/a…
1
13
1,514
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.
ai21.com/blog/verb-ai-case-s…
1
3
12
1,434
Our Co-Founder & Co-CEO, @yshoham will participate in @EntreeCap's panel discussion on #GenerativeAI. You are welcome to join!
1
13
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 @NVIDIAAI
2
2
14
1,036
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.
ai21.com/blog/zero-to-produc…
4
12
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. ai21.com/auxiliary-tuning
#NLProc #DeepLearning
3
12
Our friends at Verb have launched their AI-enhanced writing apps for authors. It's been a pleasure working with their wonderful team! 🥳
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 👇🧵
11
Since then, the space has exploded:
✅ @AI21Labs → Jamba & Jamba 1.5
✅ @NVIDIAAI → MambaVision, Nemotron-H
✅ @MistralAI → Codestral Mamba
✅ @togethercompute → Mamba-Llama
✅ @IBMResearch → Bamba
✅ @TencentGlobal → Hunyuan TurboS
✅ @MSFTResearch → Phi-4-mini-flash-reasoning
1
13
1,062
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.
1
12
962
Spotted! Our awesome Co-CEO @origoshen on a billboard (!!!) in Tel Aviv. Thanks @Google - we ❤️ you too.
2
11
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.
2
1
11
2,901
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: ai21.com/blog/ubisoft-case-s…
#gamedevelopment #AI #llm
11
844
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:
1
11
Our co-founder & co-CEO @yshoham was chosen as one of Israel's most influential leaders in deep tech! Congrats Yoav and lucky us :)
tech12.co.il/index-technolog…
1
10