Jamba2 Mini

Jamba2 Mini, published on 8 January 2026, is the second generation of the architecture that made AI21 Labs interesting. Nearly every widely used language model is a transformer, and a transformer's cost of reading text grows with the square of the input: doubling the document roughly quadruples the memory needed to keep it in the window. Jamba interleaves transformer layers with Mamba layers — a state-space design whose memory footprint grows in a straight line instead — and the practical result is that a long document stays cheap to hold in context on hardware that would choke on the same window in a pure transformer. The model totals 52 billion parameters but activates only 12 billion per token through a sparse mixture of experts, and it accepts 256,000 tokens of context. That combination is aimed squarely at the enterprise back office: contract review, long report grounding, retrieval pipelines where the retrieved material is measured in hundreds of pages rather than paragraphs. AI21's own reported strengths sit in the same place — instruction following (IFBench, IFEval, Collie), faithfulness to the supplied source (FACTS) and reliability that holds up past the 100,000-token mark — rather than in the maths-and-code leaderboards that dominate frontier model marketing. The licence is plain Apache 2.0, with no field-of-use or regional clauses, and the weights sit on Hugging Face. AI21 sells the same models as a hosted service and through cloud marketplaces, but the company's pitch is explicitly to organisations that will not send documents to someone else's API — banks, insurers, health systems — and that therefore need the model to run inside their own network. Jamba2 Mini is the middle rung of a three-model family: Jamba Large (398 billion parameters, 94 billion active, July 2025) sits above it, and Jamba 3B, small enough for on-device work, sits below. The caveat worth stating plainly is text-only. Jamba takes text in and produces text out; there is no image input, no audio and no video. In a market where the default assumption for a 2026 release is multimodality, that narrows the model to document work — which is, to be fair, exactly the work it was built for.

▸Videos

No videos yet.

▸Reviews