Jamba Reasoning 3B

What it is

Jamba Reasoning 3B is a small open reasoning model released by AI21 Labs on 8 October 2025 under the Apache 2.0 licence.^[1] It has 3 billion parameters, and AI21 says developers can download it and run it on their own devices.^[1] That is the point of the release: a reasoning model that does not have to be reached through a remote service.

A hybrid of attention and state space

The model card describes the architecture as a hybrid that combines Transformer attention with Mamba, a state-space model.^[2] The two mechanisms handle a sequence differently. Attention compares each token with the tokens around it and has to keep a cache of key-value pairs that grows with the input, which is what makes long contexts expensive in memory. A state-space model such as Mamba instead carries a compressed running state forward as it reads, so it does not need to store a large attention cache.

Jamba Reasoning 3B has 28 layers, of which 26 are Mamba layers and 2 are attention layers.^[2] The attention layers use multi-query attention with 20 query heads and one key-value head.^[2] The model card says the Mamba layers let the model process inputs without storing massive attention caches, which is how it scales to 256K tokens.^[2]

AI21 gives the context window as 256k tokens and says the model can handle up to 1M tokens; it also puts the KV cache at 8 times smaller than that of a vanilla Transformer.^[1]

Speed and benchmark scores

AI21 reported about 40 tokens per second on an M3 MacBook Pro at a 32k-token context, and named phones, Macs and PCs as the devices it is meant for.^[1]

The model card lists three scores: 61.0% on MMLU-Pro, 6.0% on Humanity's Last Exam and 52.0% on IFBench. The same card gives 70% / 5.1% / 33% for Qwen 3 4B and 42.0% / 5.2% / 28.0% for Gemma 3 4B as comparisons.^[2] Set side by side, the numbers put Jamba Reasoning 3B below Qwen 3 4B on MMLU-Pro and well above both models on IFBench, above Gemma 3 4B on MMLU-Pro, and close to the other two on Humanity's Last Exam, where all three score in the low single digits.

How it is distributed

The model is distributed through Hugging Face, Kaggle, LM Studio and llama.cpp, including GGUF quantised versions.^[1] The model card points to a separate GGUF model card for running the model locally.^[2]

Where it sits

Jamba Reasoning 3B belongs to the line of small open models judged by how well they run on ordinary hardware rather than in a data centre. Its distinguishing engineering choices are the Mamba-heavy layer stack, the long context window and the reduced cache, which together make local execution at a 32k-token context practical on a laptop.^[1][2] The published scores also show the trade-off that comes with the small size: on instruction following the 3B model sits above the two 4B models listed, while on the MMLU-Pro knowledge benchmark it trails Qwen 3 4B.^[2] Releasing the weights under Apache 2.0 and shipping quantised GGUF versions means the model can be tried without a hosted account, which is what the release is built around.^[1]