Jamba 1.6

Launched March 2025, Jamba 1.6 is AI21 Labs' flagship open model for private enterprise deployment, built on the company's hybrid SSM-Transformer architecture that interleaves Mamba state-space layers with traditional attention. The hybrid design gives it linear-time long-context processing: a 256,000-token context window with inference up to 2.5x faster than comparable Transformer-only models, and a Mixture-of-Experts structure (94B active / 398B total in Large) that keeps per-token cost down.

AI21 positioned 1.6 as the accuracy and speed leader among open enterprise models, claiming wins over Mistral, Meta, and Cohere open models on Arena Hard, RAG, and long-context QA, plus a 26-point gain in data classification over Jamba 1.5. It is validated for on-premise deployment (including HPE Private Cloud AI) and ships under the Jamba Open Model License.

Limitations: the Jamba Open Model License is not OSI-approved open source — review its commercial and redistribution terms carefully. As an enterprise-focused model it sees less community fine-tuning and tooling than Llama or Qwen, and the 1.7/Mini refresh cycle means versioning matters for reproducibility.

Vendor-reported: leads open models on Arena Hard, RAG, and long-context QA; +26pp data classification over 1.5