AI21 Labs
Hybrid Mamba-Transformer architecture with industry-leading long-context efficiency.
About
About AI21 Labs
AI21 Labs, founded in Tel Aviv in 2017, pioneered the Jamba architecture — a hybrid combining Mamba state-space models with Transformer attention layers. This allows Jamba to process 256K tokens at significantly lower memory overhead than pure-attention models. Jamba 1.5 Large delivers GPT-4 class quality with exceptional throughput, making it the standout choice for long-document processing pipelines.
Key highlights
Jamba is the first widely-available Mamba+Transformer hybrid architecture
256K context window with lower memory usage than comparable pure-attention models
Jamba 1.5 Mini offers 256K context at $0.20/M — exceptional value for long documents
Strong multilingual support and structured output generation
Available on AWS Marketplace, Azure AI Studio, and the AI21 API
Models
2 models tracked
Context windows, pricing, capabilities, and API identifiers — everything you need to choose the right model for your workload.
Jamba 1.5 Large
2024
AI21's flagship model using the Jamba hybrid architecture. Processes 256K tokens efficiently by combining Mamba state-space layers with selective Transformer attention. GPT-4 class quality.
- 256K context window
- Mamba+Transformer hybrid
- Low memory overhead
- GPT-4 class quality
jamba-1.5-largeJamba 1.5 Mini
2024
A cost-efficient Jamba model retaining the full 256K context window. At $0.20/M input, it is among the cheapest models capable of processing book-length documents.
- 256K context at ultra-low cost
- Very high throughput
- Efficient long-context
- Cost-effective
jamba-1.5-miniUse Cases
What AI21 Labs is best for
Each provider has specialisms shaped by their architecture decisions, training data, and API design. Here is where AI21 Labs performs best.
Get started
Start building with AI21 Labs
Visit the official AI21 Labs platform to access API keys, documentation, and pricing details.