Jamba

Jamba combines Transformer attention layers with Mamba State Space Model (SSM) layers in a hybrid architecture that enables efficient processing of 256K token context windows. The architecture interleaves attention and SSM layers to balance computational efficiency with semantic understanding, allowing the model to process extended documents (financial records, contracts, knowledge bases) without the quadratic memory scaling of pure Transformer models. This hybrid approach enables 'up to 30% more text per token' efficiency compared to standard tokenizers while maintaining strong performance on reasoning and generation tasks.

Jamba2 3B and Jamba Mini variants are optimized for on-device deployment with 3 billion parameters, enabling inference on edge devices, mobile hardware, and resource-constrained environments without cloud API calls. The compact parameter count combined with the hybrid Mamba-Transformer architecture reduces memory footprint and latency compared to larger models, while maintaining performance on agentic workflows and reasoning tasks. Models are available as open-source downloads from Hugging Face in formats suitable for local deployment.

Jamba API supports batch processing for high-volume inference workloads, enabling cost optimization through deferred execution and bulk token pricing. Batch processing allows applications to submit multiple requests for asynchronous processing, reducing per-token costs and enabling cost-effective processing of large document collections or periodic analysis tasks. This is particularly valuable for long-context workloads where per-token costs are significant.

AI21 offers custom enterprise plans for large-volume deployments, including volume discounts on per-token pricing, premium rate limits, private cloud hosting, and dedicated technical support. Enterprise customers can negotiate custom SLAs, priority access to new models, and domain-specific fine-tuning. This enables organizations to optimize costs at scale and receive dedicated support for production deployments.

Jamba Reasoning 3B variant is specifically tuned for complex reasoning tasks while maintaining the 256K context window, enabling multi-step logical inference over extended documents and conversation histories. The model uses chain-of-thought patterns and is optimized for 'record latency' on reasoning workloads, making it suitable for enterprise decision-making systems that require both speed and accuracy. Available via AI21 Studio API with usage-based pricing ($0.2/1M input, $0.4/1M output tokens for Mini variant).

Jamba models are accessible via AI21 Studio cloud API with usage-based pay-as-you-go pricing, supporting multiple model variants (Mini, Large, Reasoning 3B) with transparent per-token costs. The API provides REST endpoints for text generation with configurable parameters (temperature, max tokens, top-p sampling) and supports batch processing for cost optimization. Pricing ranges from $0.2/1M input tokens (Mini) to $2/1M input tokens (Large), with output token pricing 2-4x higher than input.

Jamba models are available as open-source downloads from Hugging Face, enabling self-hosted deployment without API dependencies or cloud costs. Models are distributed in standard formats compatible with inference frameworks (vLLM, Ollama, llama.cpp, etc.) and support both CPU and GPU inference. The open-source availability enables fine-tuning, quantization, and custom optimization for specific use cases, with no licensing restrictions documented for commercial use.

Jamba offers multiple model variants (Mini, Large, Reasoning 3B, 2 3B) optimized for different cost-performance tradeoffs, enabling builders to select the appropriate model for their use case without over-provisioning. Mini variants prioritize efficiency and cost ($0.2/1M input tokens), while Large variants provide maximum capability ($2/1M input tokens), and Reasoning 3B targets reasoning workloads. All variants share the 256K context window and hybrid architecture, allowing seamless switching based on workload requirements.

Jamba is optimized for enterprise verticals including finance, healthcare, defense, technology, and manufacturing, with specific tuning for domain-specific tasks like financial analysis, contract review, and compliance checking. The 256K context window and reasoning capabilities enable processing of domain-specific documents (financial reports, medical records, contracts) without truncation. AI21 offers custom enterprise plans with domain-specific fine-tuning and dedicated support for vertical-specific deployments.

Jamba supports on-premises and sovereign AI deployment for organizations with data residency, security, or geopolitical requirements. Models are available as open-source downloads for self-hosting, and AI21 offers custom enterprise plans with private cloud hosting, dedicated infrastructure, and compliance certifications. This enables organizations to maintain full data control and meet regulatory requirements (GDPR, HIPAA, national security) without sending data to external cloud providers.

Jamba achieves 'up to 30% more text per token' efficiency compared to standard tokenizers through optimized tokenization, reducing the number of tokens required to represent the same text. This efficiency gain directly reduces API costs (fewer tokens billed) and increases effective context window capacity (more text fits within 256K token limit). The tokenization improvement applies across all model variants and deployment methods (API and self-hosted).

Jamba2 3B is specifically optimized for agentic workflows, enabling the model to plan multi-step tasks, call external tools, and maintain state across interactions. The model supports function calling and tool integration patterns required for autonomous agents, with the compact 3B parameter size enabling on-device agent deployment. The 256K context window allows agents to maintain full conversation history and tool execution logs without truncation.