AI21 Jamba 1.5

Generates text using a hybrid architecture that interleaves Mamba structured state space (SSS) layers with Transformer attention layers, enabling linear-time sequence processing instead of quadratic complexity. The Mamba layers maintain recurrent state across 256K token contexts while Transformer layers provide attention-based refinement, allowing efficient inference on documents up to 256K tokens without the memory explosion of pure Transformer models. This architecture enables processing of entire books, legal contracts, or multi-document datasets in a single forward pass.

Provides instruction-following and conversational capabilities through fine-tuned Chat and Instruct variants optimized for enterprise use cases across Finance, Tech, Defense, Healthcare, and Manufacturing domains. The model follows natural language instructions with context awareness maintained across the 256K token window, enabling multi-turn conversations that reference earlier context without degradation. Deployed via AI21 Studio API with usage-based pricing or self-hosted on customer infrastructure.

Jamba models are released as open-source with weights available on Hugging Face, enabling community contributions, research, and custom deployments. The open-source approach allows researchers to study the hybrid Mamba-Transformer architecture, contribute improvements, and build upon the models. Community members can create optimized inference implementations, fine-tuning guides, and domain-specific adaptations without licensing restrictions.

Achieves inference efficiency through the Mamba SSS architecture which eliminates the quadratic memory scaling of Transformer self-attention, reducing GPU VRAM requirements compared to models of similar capability. The hybrid design balances efficiency gains from Mamba layers with quality preservation from Transformer layers, enabling deployment on resource-constrained infrastructure. Supports both API-based inference via AI21 Studio and self-hosted deployment with configurable hardware.

Provides hosted inference via AI21 Studio API with transparent usage-based pricing ($0.2-$0.4/1M tokens for Mini, $2-$8/1M tokens for Large) and free trial credits ($10 for 3 months, no credit card required). Supports both Jamba Mini (12B active) and Large (94B active) variants with identical API interface, enabling cost-optimization by selecting appropriate model size per use case. Integrates with standard HTTP/REST patterns and SDKs for Python and other languages.

Enables deployment of Jamba models on customer-controlled infrastructure (on-premises or private cloud) via model downloads from Hugging Face and integration with standard inference frameworks. Supports deployment through 'trusted technology partners' (partners not named in documentation) and custom cloud deployments. Provides full model control, data privacy, and elimination of API latency at the cost of infrastructure management and operational complexity.

Leverages the 256K context window to simultaneously process and synthesize information across multiple related documents (financial reports, research papers, contracts, etc.) in a single inference pass. The hybrid Mamba-Transformer architecture maintains coherent understanding across document boundaries while the linear-time complexity enables processing of dozens of documents without memory explosion. Enables cross-document reasoning, contradiction detection, and synthesis without lossy summarization or chunking.

Claims to achieve up to 30% more text per token than competing providers through optimized tokenization, reducing the effective cost of long-context processing and enabling more content to fit within the 256K token window. The tokenization approach is not documented, but the claim suggests more efficient encoding of natural language compared to standard BPE or SentencePiece tokenizers used by other models.

Distributes Jamba model weights via Hugging Face Model Hub as open-source models, enabling free download, inspection, and modification without licensing restrictions. Both Mini (12B active/52B total) and Large (94B active/398B total) variants are available, allowing developers to use, fine-tune, and redistribute models under open-source terms. Supports integration with standard Hugging Face tooling (transformers library, model cards, community discussions).

Jamba Mini uses only 12B active parameters out of 52B total parameters through sparse activation patterns, and Jamba Large uses 94B active per 398B total, enabling inference with reduced computational cost compared to dense models of equivalent quality. The hybrid architecture with Mamba layers contributes to this efficiency by avoiding the dense attention computations of pure Transformers. This sparsity pattern is similar to mixture-of-experts approaches but implemented through the Mamba-Transformer hybrid design.

Positions Jamba for enterprise use across Finance, Tech, Defense, Healthcare, and Manufacturing domains with claims of domain-specific optimization, though no domain-specific model variants or fine-tuning details are documented. The 256K context window and efficient inference enable deployment in enterprise environments with large document volumes and strict latency/privacy requirements. Available through 'trusted technology partners' for cloud deployment (partners not named).

AI21 Jamba 1.5 is a cutting-edge language model designed for long document understanding and multi-document tasks, featuring a massive 256K context window and efficient inference.