Enterprise context windows and the AI21 Jamba strategy

The architecture of long-context efficiency

Google Gemini 2.5 Pro supports a 1 million token context window, which allows users to process 50,000 lines of code or 8 average-length English novels. AI21 Labs provides a different solution for businesses with its Jamba model family, which uses a 256K token context window. The Jamba architecture combines Mamba state space models with Transformer layers to handle the trade-off between speed, memory, and quality. This hybrid design uses Mamba layers for short-range dependencies and Transformer layers for long-range dependencies. While pure Transformer models face quadratic memory scaling, Jamba’s Mamba components use linear scaling to manage sequence length more efficiently. This approach helps the model avoid the high costs and latency issues that typically plague long-context processing. The Jamba-Instruct model is the aligned version of the base model, and it includes safety guardrails and enhanced command comprehension for immediate enterprise use. This model can intake the equivalent of a 400 page novel or every financial filing for a company in a single fiscal year. AI21 Labs was the first independent startup to surpass OpenAI’s GPT-3. The company started in Tel Aviv, where it utilizes a deep talent pool in applied mathematics and computer science. Shoham and Goshen met in an Israeli non-profit project that democratized programming skills.

Performance and readiness

Jamba 1.5 Mini achieves over 150 output tokens per second for prompts of 10,000 tokens. The RULER benchmark, which was released by NVIDIA to evaluate models on multi-step reasoning tasks including retrieval, multi-hop tracing, aggregation, and question answering, shows that Jamba hits an 85% pass rate for its effective context length. Many competitors claim massive context windows but fail to achieve useful performance once they exceed certain limits. Jamba 1.5 Large outperforms competitors like Llama 3.1 70B and Mistral Large 2 in latency tests. You probably understand that increasing context length usually forces a choice between accuracy and serving costs. Because of this hybrid structure, the model handles tasks with low latency. In tests with 10,000 tokens, Jamba 1.5 Mini ranked as the fastest model among its size class. One major drawback is that most long-context models struggle to actually perform useful tasks with the extra tokens they claim to support. A single 80GB GPU can run a 140,000-token context window with Jamba. This capability allows companies to build RAG pipelines where the model retrieves snippets from millions of documents to provide accurate answers. This efficiency helps retail companies generate product descriptions or helps customer service agents through grounded question answering.

Enterprise deployment and scale

AI21 Labs reached a $1.4 billion valuation following a $300 million Series D in May 2025. Investors in this round included Google and Nvidia. The company focuses on enterprise-level customers that require high-performance, secure LLMs for legal, financial, and healthcare workflows. AI21 distributes its models through Amazon Web Services, Google Cloud, Microsoft Azure, Snowflake, Databricks, and NVIDIA. Developers can access Jamba 1.5 models via Azure AI as Models-as-a-Service, which means they pay for usage without managing infrastructure. The input cost for Jamba 1.5 Mini is $0.0002 per 1,000 tokens, while Jamba 1.5 Large costs $0.002 per 1,000 input tokens. This pricing structure targets companies that need to build RAG pipelines where the model retrieves snippets from millions of documents to provide accurate answers. The company’s revenue crossed $100 million in annual recurring revenue by early 2026. AI21 Labs also provides Jamba 1.6 as an NVIDIA NIM microservice for private deployment in regulated environments. This allows developers to deploy Jamba 1.6 in under five minutes using standardized APIs. The Jamba 1.6 model is suitable for financial services needing loan term sheet generation or healthcare firms needing research assistants. AI21 Labs previously launched Wordtune, an AI writing assistant that has over 10 million users. The company also provides access to the Jurassic-2 model family through its developer platform. Can enterprises rely on smaller, more efficient context windows when the leading competitors provide millions of tokens?