As AI consultants based in Rosenheim, Germany, we recommend AI21 Jamba for enterprises that need to process very long documents. AI21 Labs is an Israeli AI company that developed a hybrid architecture combining Mamba (a state space model) and Transformer for Jamba. With a 256k token context window, entire books can be analyzed at once.
The relevant model family today is Jamba2 (January 2026, Apache 2.0, open weights on Hugging Face). In May 2026, AI21 shifted its strategic focus to the Maestro orchestration platform for enterprise AI agents, laid off more than 60% of its workforce and is scaling back the sale of Jamba models as a standalone API product. The open weights remain available under Apache 2.0 and can be self-hosted. This overview reflects the state as of October 2026.
Current models at a glance
| Model | Release | License | Status |
|---|---|---|---|
| Jamba2 Mini | January 2026 | Apache 2.0 | current, recommended |
| Jamba2 3B | January 2026 | Apache 2.0 | current |
| Jamba Reasoning 3B | October 2025 | Apache 2.0 | current |
| Jamba Large 1.7 | July 2025 | Jamba Open Model License | current |
| Jamba 1.6 (Large/Mini) | March 2025 | Jamba Open Model License | deprecated |
| Jamba 1.5 Large / Mini | August 2024 | Jamba Open Model License | deprecated, Legacy on Bedrock |
Architecture: Mamba-Transformer hybrid
Jamba combines Mamba layers (efficient processing of long sequences), Transformer layers (precise attention for detail) and mixture-of-experts for efficiency. According to AI21, this yields faster inference at long contexts, lower memory requirements and linear rather than quadratic scaling.
256,000 tokens correspond to roughly 640 pages of text: entire books, complete codebases or extensive legal contracts can be processed in one pass.
Hardware requirements (self-hosted)
| Model | VRAM | Recommended GPU |
|---|---|---|
| Jamba2 Mini | 48 GB | A100 80GB |
| Jamba 1.5 Large | 160 GB | Multi-A100 |
| Jamba 1.5 Mini | 24 GB | RTX 4090 |
EU availability and GDPR
According to the AWS model overview, Amazon Bedrock carries only the older Jamba 1.5 models from AI21 (in-Region only in us-east-1, lifecycle status “Legacy”, AWS EOL date November 26, 2026) – never in Frankfurt or any other EU region. The newer Jamba2 and Jamba 1.6/1.7 models are not listed on Bedrock. Genuine EU availability therefore exists only through self-hosting the open Jamba2 weights on your own EU infrastructure, which also means full data control.
Model lifecycle
- Jamba 1.5 (Large/Mini): under the Jamba Open Model License; Legacy on Bedrock with EOL on November 26, 2026, and since August 26, 2026 in the ‘public extended access’ phase, in which higher prices may apply per AWS.
- Jamba 1.6: deprecated on the API as of August 3, 2025 per AI21’s documentation, superseded by Jamba Large 1.7 and Jamba2.
- Jamba Mini 1.7: deprecated on the API as of February 1, 2026 per AI21’s documentation; the alias
jamba-minipoints to Jamba2 Mini (jamba-mini-2-2026-01). - AI21 API: the sale of Jamba models as a standalone API product is being scaled back; evaluate alternative providers for new cloud-API projects.
Integration with CompanyGPT
AI21 Jamba can be integrated into CompanyGPT – ideal for companies with extensive document collections.
Our recommendation
AI21 Jamba2 Mini as an open-weights model under Apache 2.0 is our recommendation for self-hosted document analysis of very long texts. If you regularly work with very long documents (legal contracts, book manuscripts, code reviews) and need full data sovereignty, Jamba2 on your own infrastructure is a good choice.
Given AI21’s strategic shift to the Maestro platform, we advise caution for API-centric new projects: rely on self-hosting the open weights or evaluate alternatives. For general cloud-API applications without a special focus on long contexts, we recommend Google Gemini (up to 1M context) or OpenAI GPT.