AI21 Labs releases Jamba Reasoning 3B for on-device use

Background

AI21 Labs was extending its Jamba family of language models. The model's design reflected a practical concern: keeping memory use low when a model reads long documents.

What happened

On 8 October 2025 AI21 Labs released Jamba Reasoning 3B, a 3-billion-parameter open reasoning model ^[1]. A reasoning model works through a problem step by step before it answers. The model was released under the Apache 2.0 licence, a permissive open-source licence ^[1]. AI21 targeted phones, laptops and PCs with it, citing a speed of about 40 tokens per second on an M3 MacBook Pro ^[1]. A token is a small piece of text, so that figure measures how fast the model writes out words on that laptop.

Key details

The model's structure combines 26 Mamba layers with 2 attention layers ^[2]. Attention layers let each part of a text look at every other part, which gives strong results but uses more memory as inputs grow. Mamba layers handle long sequences with less memory. AI21 chose this mix to keep memory use low on long inputs ^[2].

The model has a 256k-token context window, so it can read a long document or conversation in one pass ^[1].

The model card reports scores of 61.0% on MMLU-Pro and 52.0% on IFBench ^[2]. MMLU-Pro is a broad test of knowledge and reasoning, and IFBench measures how well a model follows instructions. These are the figures the card gives for this model, and they can be compared with the scores the card lists for other small models.

Impact

The release shows AI21 putting a reasoning model on hardware that ordinary users own, not only on servers. Its open licence means that anyone can download it and test it on their own computer ^[1]. These performance figures are the company's own.