Jamba 1.6

A hybrid SSM-Transformer family

Jamba 1.6 is a family of foundation models released by AI21 Labs in March 2025 for private enterprise deployment ^[2]. AI21 describes the family as hybrid SSM-Transformer instruction-following foundation models ^[3]. The name points at the architecture. Most of the network is built from state space model (SSM) layers, which read a sequence step by step and carry a compressed internal state forward as they go; the rest are ordinary attention layers, in which every position can look directly at earlier positions. A state space layer stays cheap over very long inputs because the memory it carries does not grow with the length of the text, while attention is expensive but precise at pulling out a particular earlier token. Stacking the two gives a model that can keep a long document in view without paying the full cost of attention across all of it.

Two sizes

The family ships in two sizes. Jamba Large 1.6 has 398B total parameters, of which 94B are active; Jamba Mini 1.6 has 52B total and 12B active ^[1]. Total parameters are everything the model stores, while active parameters are the ones actually used in computing a single token, so a model can be much larger than the part of it engaged at any one moment. That gap is what makes a model of this size practical to serve.

Both models accept a 256k-token context and support English, Spanish, French, Portuguese, Italian, Dutch, German, Arabic and Hebrew ^[1]. A context window of that size is aimed at inputs that do not fit in a prompt-sized window: several documents at once, long conversations, large code files.

Benchmarks

AI21 published its own comparison of Jamba Large 1.6 against Mistral Large 2, Llama 3.3 70B and Command R+ ^[1]. On Arena Hard, a set of difficult prompts scored by comparing model outputs, Jamba Large 1.6 scores 76.5 against 71.55, 65.77 and 33.05 for the three others. On CRAG, which tests answering questions with retrieved documents, it scores 78.2 against 67, 61.7 and 49.4. On LongBench, a long-context suite, it scores 38.8 against 34.4, 21.7 and 18.9 ^[1]. These are AI21's reported figures; benchmark scores shift with prompt formats and settings, so a table assembled by one vendor should be read as that vendor's own measurement rather than a definitive ranking.

For the smaller model the Jamba Mini 1.6 model card reports Arena Hard 51.2, CRAG 76.2, FinanceBench (FullDoc) 45.4, HELMET LongQA 46.9 and LongBench 32 ^[3]. FinanceBench asks questions over financial filings, and HELMET LongQA tests question answering over long inputs; both are the kind of task the long context and the hybrid architecture are meant to serve.

Release and licensing

The weights were released under the Jamba Open Model License as gated downloads on Hugging Face ^[1]: the models can be downloaded and run, but access goes through an approval step and use is governed by AI21's own licence rather than a fully unrestricted one. The family launched in the same month as Maestro, AI21's planning and orchestration system ^[2], which places the models alongside a layer for coordinating tasks rather than as standalone chat products.

Where it sits in the line

Jamba 1.6 was succeeded by Jamba 1.7 ^[1], and the Jamba Mini 1.6 model card points readers to Jamba Mini 1.7 as a newer version ^[3]. The 1.6 release is therefore a snapshot of a fast-moving family: two sizes, a 256k-token context, an open-but-gated licence, and vendor-reported results positioned against larger-name competitors of the moment.