Jamba 1.6 Large

Jamba 1.6 is a hybrid SSM-Transformer foundation model that redefines what open models can offer enterprises. With a 256K context window, 94B active parameters, and ExpertsInt8 quantization, it delivers industry-leading performance on long-context and grounded reasoning tasks—without compromising speed or deployability. On LongBench, it scores 38.8, outperforming Mistral Large 2 (34.4) and LLaMA 3.3 70B (21.7). On CRAG, which evaluates citation-grounded QA, it leads with 78.2, and its 76.5 Arena-Hard score signals top-tier instruction following. Remarkably, this performance holds up in real-world latency-constrained settings, with up to 2.5× faster inference than comparable models, making it practical for use cases like document synthesis, enterprise chatbots, and RAG pipelines at scale. AI21 is positioning Jamba not just as a benchmark leader but as a deployment-ready, private-by-default alternative to GPT-4-tier closed models. Customers like Fnac and Educa EdTech report significant gains in output precision, latency, and retrieval accuracy—some even switching from Jamba Large 1.5 to Jamba Mini 1.6 without quality loss. Jamba’s full-stack support for structured outputs, tool use, and fine-tuning (via qLoRA + FSDP) makes it unusually accessible for enterprise customization. Unlike other open models, Jamba is architected for private VPC or on-prem deployment, allowing regulated industries to maintain full control of data while leveraging state-of-the-art model capabilities. In a fast-maturing open ecosystem, Jamba 1.6 is setting a new bar for what open actually enables.