Jamba 1.6 Released

AI21 released Jamba 1.6, emphasizing the performance advantage of its open model family for enterprise deployment. Jamba Large 1.6 surpassed Mistral Large 2, Llama 3.3 70B, and Command R+ in quality, and Jamba Mini 1.6 also delivered higher performance than Ministral 8B, Llama 3.1 8B, and Command R7B, the company said.

The core is 256K context and a hybrid SSM-Transformer architecture. AI21 explained that Jamba 1.6 shows strength in RAG and long-context question answering, minimizing accuracy degradation even as context grows longer. The comparison data was based on Artificial Analysis speed scores and the LongBench and Arena Hard leaderboards.

Deployment options are also flexible. Beyond AI21 Studio, the model can be downloaded from Hugging Face and run on-premise or in-VPC, with additional deployment options announced as forthcoming. The company argued that this structure, which prevents sensitive data from being exposed to the model vendor, reduces the need to choose between security and performance.

Real-world use cases were also presented.

• Fnac: With Jamba 1.6 Mini, output quality improved by 26% and latency was reduced by about 40%
• Educa Edtech: Achieved over 90% retrieval accuracy and citation reliability in grounded QA
• A digital banking customer: In internal testing, Jamba Mini 1.6 recorded 21% higher precision than the previous model and showed quality on par with GPT-4o, the company explained

Finally, a new Batch API was introduced, enabling asynchronous processing of large-volume requests. In Fnac's testing, wait times for tens of thousands of requests dropped from several hours to under 1 hour, significantly improving efficiency for tasks such as large-scale document processing and product description review, the company said.