Jamba2 3B

Jamba2 3B is the smallest member of the Jamba family, published by AI21 Labs on 8 January 2026 alongside Jamba2 Mini. The vendor positions it for work that never leaves the device: phones, laptops and desktop machines, where a model has to fit in ordinary consumer memory rather than in a data centre. The architecture is what makes the size interesting. Jamba interleaves Mamba layers, which are state-space models that read text in a single pass at constant memory cost, with a small number of ordinary attention layers. The published weight configuration shows how sparingly attention is used here: of 28 layers, an attention layer appears only once every fourteen, so the overwhelming majority of the network is Mamba. Unlike Jamba2 Mini, this model has no mixture of experts — every parameter is used on every token. That is a deliberate trade for edge hardware, where the memory needed to hold unused experts is exactly what a phone does not have. The practical consequence is the context window. A transformer's memory cost for a conversation grows with the square of its length, which is why small models are usually capped at short contexts; Jamba2 3B keeps 256,000 tokens, the same figure AI21 quotes for models more than a hundred times its size. The vendor reports category-leading results on instruction-following benchmarks (IFBench, IFEval, Collie) and on the FACTS grounding benchmark, but publishes those results only as charts without numbers, so no score is repeated here. AI21 built the Jamba2 generation on the pre-training of Jamba 1.5, then added mid-training on 500 billion curated tokens weighted towards mathematics, code and long documents, a separate phase tuning the Mamba layers for long contexts, supervised fine-tuning, DPO and several rounds of reinforcement learning. One limitation is worth stating plainly: unlike the rest of the family, Jamba2 3B has no API endpoint at AI21 — the vendor's own model table lists it as not available. It exists as downloadable weights on Hugging Face, and nowhere else. The cloud marketplaces are further behind still: as of this writing Amazon Bedrock and Azure carry only the 1.5 generation, Google Cloud 1.6 and Kaggle 1.7.

▸Videos

No videos yet.

▸Reviews