News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
Jamba2 Mini

Jamba2 Mini

A small open-weights SSM-Transformer model for enterprise work

Jamba2 Mini is a small language model from AI21 Labs, released as open weights under Apache 2.0 with a 256K context window. It mixes a state-space design with attention and is aimed at enterprise tasks such as question answering and agent stacks, where steady, grounded output matters more than heavy reasoning.

Share on WhatsApp

  • 52B total parameters, 12B active per token
  • 256K token context window
  • Apache 2.0 licence, weights on Hugging Face
  • SSM-Transformer hybrid with Mamba layers
  • Mid-training on 500B curated tokens

What it is

Jamba2 Mini is a small language model that AI21 Labs publishes as open source for enterprise use. It carries 52 billion parameters in total but draws on only 12 billion of them for each token it produces, which keeps the running cost down. The model answers questions and follows instructions rather than doing long chains of reasoning. Its context window spans 256K tokens, enough to hold technical manuals, research papers or whole knowledge bases at once. AI21 Labs positions it for production agent stacks, where outputs need to stay consistent and tied to the documents given to it.

How it came about

The Jamba2 family grew out of the earlier Jamba 1.5 line. AI21 Labs says the new models were built on top of Jamba 1.5 pre-training and then pushed through a longer post-training pipeline. The company wrote a release blog post introducing the family, and put the Mini variant on Hugging Face under the Apache 2.0 licence. Exact release dates are not given here.

Inside the model

The model uses an SSM-Transformer design, mixing state-space layers with attention so that memory use stays low even at very long contexts. Training began from Jamba 1.5 pre-training, then a mid-training stage on 500B curated tokens weighted towards maths, code, high-quality web text and long documents. A state passing stage tuned the Mamba layers for longer contexts. Cold-start supervised fine-tuning taught instruction following and reasoning, followed by DPO. The last stages were several rounds of on-policy reinforcement learning, moving from short contexts with verifiable rewards to longer ones with mixed verifiable and model-based rewards.

What it can do

AI21 Labs reports strong results on instruction following and grounding, naming IFBench, IFEval, Collie and FACTS among the benchmarks. In blind side-by-side tests on 100 real-world enterprise prompts, human raters gave it statistically significant wins on output quality and factuality against Ministral3 14B. Ratings covered factuality, style, constraint adherence, instruction following and helpfulness. The company says it holds up at contexts beyond 100K tokens.

How to use it

Weights are downloadable from Hugging Face under the Apache 2.0 licence, which allows commercial use. The model card gives setup guides for vLLM (version 0.12.0 or higher) and for Transformers, with flash-attention and causal-conv1d among the dependencies. AI21 Labs also points users to its own documentation site.

Questions people ask about Jamba2 Mini

What kind of model is Jamba2 Mini?

It is a small, general-purpose language model from AI21 Labs built for enterprise work. It answers questions and follows instructions over long documents rather than doing extended reasoning, and it runs on an SSM-Transformer design that keeps memory use low.

How large is it?

The model has 52 billion parameters in total, but only 12 billion are active for each token it generates. That combination is meant to give the quality of a larger model at a lower running cost.

Can I use it commercially?

Yes. The weights are released under the Apache 2.0 licence, which AI21 Labs describes as fully open source for commercial use. They can be downloaded from Hugging Face and run with vLLM or Transformers.

How much text can it handle at once?

Its context window is 256K tokens, so it can take in long inputs such as technical manuals, research papers or knowledge bases. AI21 Labs says performance holds up at contexts of 100K tokens and beyond.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Jamba2 Mini yet, so the facts on this page come from the sources listed above: huggingface.co/ai21labs/AI21-Jamba2-Mini. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.