AI21 Labs

Jamba2

AI21 Labs

Jamba2 is an AI21 language-model release. Jamba2 is an enterprise-focused release emphasizing grounded answers and instruction adherence with concise outputs. Dense 3B and sparse Mini tiers share the hybrid design and 256K context. Both release under Apache 2.0, expanding commercial weight access relative to the custom-licensed 1.x instruction models. Public weights and AI21 Studio support local and hosted deployment. Long context does not guarantee complete retrieval or factual correctness, and deployed throughput depends on hardware, precision, and prompt length.

Jamba Reasoning 3B

AI21 Labs

Jamba Reasoning 3B is an AI21 language-model release. A compact reasoning specialization for local devices and efficient long-context agents. Post-training combines supervised, preference, and reinforcement-learning methods to support mathematics, code, instruction following, and general reasoning. The documented default context is 256K; claims of extended 1M processing are not treated as a uniform deployment guarantee. Long context does not guarantee complete retrieval or factual correctness, and deployed throughput depends on hardware, precision, and prompt length.

Jamba 1.7

AI21 Labs

Jamba 1.7 is an AI21 language-model release. Jamba 1.7 improves grounded answers, instruction following, and optimized serving for self-hosted enterprise work. AI21’s July 3 changelog establishes release, avoiding later dates when third-party endpoints adopted the model. The large tier has 398B total parameters and 94B active; smaller Mini has a different deployment footprint. Long context does not guarantee complete retrieval or factual correctness, and deployed throughput depends on hardware, precision, and prompt length.

Jamba 1.6

AI21 Labs

Jamba 1.6 is an AI21 language-model release. The 1.6 update improves enterprise question answering, retrieval grounding, multilingual instruction following, and long-context quality. It is available through AI21 Studio and downloadable private deployments. The provider changelog explicitly records March 6 even though the containing update is dated March 22. Long context does not guarantee complete retrieval or factual correctness, and deployed throughput depends on hardware, precision, and prompt length.

Jamba 1.5

AI21 Labs

Jamba 1.5 is an AI21 language-model release. Jamba 1.5 introduces instruction-tuned Mini and Large tiers for grounded enterprise conversation and tool workflows. It retains the hybrid state-space/attention structure and long context, while downloadable weights use AI21’s custom license rather than the original base model’s Apache license. Long context does not guarantee complete retrieval or factual correctness, and deployed throughput depends on hardware, precision, and prompt length.

Jamba

AI21 Labs

Jamba is an AI21 language-model release. The original Jamba combines attention and Mamba state-space layers with sparse experts. It provides a 256K context with a smaller attention cache than a pure Transformer, aiming at efficient long-document language modeling. The original checkpoint is a base completion model; instruction-tuned derivatives and later versions have different terms. Long context does not guarantee complete retrieval or factual correctness, and deployed throughput depends on hardware, precision, and prompt length.

Jurassic-2

AI21 Labs

Jurassic-2 is a proprietary AI21 language-model family. The second Jurassic generation improves instruction following, multilingual support, and API responsiveness, with tiered models for enterprise text applications. It was released with specialist APIs for tasks such as paraphrasing and summarization. The model’s full parameter and architecture details were not publicly specified. This is a historical proprietary release superseded by the Jamba family. No public weights are available. This record describes first publication and does not imply that the original endpoints remain supported. Generated content may be inaccurate or biased.

Jurassic-1

AI21 Labs

Jurassic-1 is a proprietary AI21 language-model family. AI21’s initial developer-accessible language-model family was offered in an open beta through AI21 Studio. It introduced a large 250,000-token vocabulary including multiword tokens, supporting text generation, question answering, and writing applications. The current official announcement is dated August 4, while contemporary reporting commonly dates wider launch to August 11. No public weights are available. This record describes first publication and does not imply that the original endpoints remain supported. Generated content may be inaccurate or biased.