MDL-3140EST.2025 · IDX.340
Language modelIn production

Jamba Reasoning 3B

AI21 Labs · Israel · 2025

AI21's first reasoning model and its most downloaded release — three billion parameters that run on a laptop, with a 256,000-token window and published benchmark figures rather than the usual unlabelled charts.

wujec.ai score

/10

Community score

no votes yet
Sign in to rate

Jamba Reasoning 3B, released on 8 October 2025, was AI21 Labs' first attempt at a model that thinks before it answers, and it was aimed at hardware nobody associates with reasoning: laptops, phones and single consumer GPUs. The vendor quotes 40 tokens per second on an M3 MacBook Pro at a 32,000-token context. The layout is the Jamba hybrid at small scale, and here the vendor states it exactly: 28 layers, of which 26 are Mamba state-space blocks and only 2 are attention layers, with 20 query heads sharing a single key-value head. That ratio is the reason a three-billion-parameter model can hold 256,000 tokens of context — attention is what forces a model to store a growing cache as a conversation lengthens, and this network barely uses it. On results the model deserves credit for publishing actual numbers, which AI21 did not do for the later Jamba2 generation. It reports MMLU-Pro 61.0%, Humanity's Last Exam 6.0% and IFBench 52.0%, against named competitors in the same table. The IFBench figure is the striking one: 52% where the closest rival in the vendor's own table reaches 33%. One qualification belongs next to the vendor's claim of outperforming the field, because it comes from that same table: on MMLU-Pro, Qwen 3 4B scores 70% against Jamba Reasoning 3B's 61%. AI21's claim rests on an average across six benchmarks, not on winning each of them. The licence needs a note too. AI21's announcement states plainly that the model is released under the Apache 2.0 licence; the model card on Hugging Face carries the Apache 2.0 label but also links a separate AI21 open-model licence. The announcement is the vendor's clearest statement, and it is what this profile follows. The model was superseded on 8 January 2026 by Jamba2 3B, which keeps the same size and context window but drops the reasoning framing in favour of instruction following and grounding. Jamba Reasoning 3B remains downloadable, including a quantised GGUF build for llama.cpp and LM Studio, and remains AI21's most downloaded recent release.

#open weights#Apache 2.0#Mamba#reasoning#on-device#long context
Official website

Videos

No videos yet.

Reviews

No reviews yet. Be the first!

Sign in to write a review