Last updated:
Jamba
AI21 Labs' Jamba models — hybrid SSM-Transformer models built for long context, now shifting to AI21's Maestro platform.
Model Facts
Latest model: Jamba2 Mini
- Provider
- AI21 Labs
- Released
- Jan 1, 2026
- Context window
- 256K tokens
- Weights
- Open
Full Model List
| Model | Released | Status | Context | Note |
|---|---|---|---|---|
| Jamba2 Mini | Jan 1, 2026 | current | 256K tokens | 52B-parameter MoE (12B active), 256K context, Apache 2.0; month approximate |
| Jamba2 3B | Jan 1, 2026 | current | – | 3B-parameter efficient model, Apache 2.0; month approximate |
| Jamba Large 1.7 | Jul 1, 2025 | legacy | 256K tokens | hybrid SSM-Transformer, 256K context; date approximate |
- Official model page:
- ai21.com
- Official documentation:
- docs.ai21.com
- Other sources:
- AI21 Jamba (innFactory)
What it is
Jamba is AI21 Labs' family of hybrid SSM-Transformer models. The Jamba2 generation arrived in early 2026: Jamba2 Mini (52B parameters, 12B active, 256K context) and Jamba2 3B, both Apache 2.0. As of May 2026 AI21 has refocused on its Maestro orchestration platform for enterprise agents and is phasing back standalone Jamba API sales, though the open weights remain available.
Where this information comes from
Details come from model trackers; exact Jamba2 release days were not confirmed. Release dates, context windows and benchmark scores are vendor-reported unless stated, and harnesses differ between labs, so treat the numbers as indicative.
Notable traits
- Long context from a hybrid architecture.
- Open weights remain downloadable.
- AI21's focus has moved to agent orchestration.