Codestral Mamba
Codestral Mamba is a 7.3B Mamba-based code model with linear-time inference, tested to 256K context, under Apache 2.0.
A non-Transformer code model. Mamba's state-space layers give linear-time inference and, in theory, unbounded sequence length. Mistral says it performs comparably to Transformer code models of similar size.
- Date
- Tuesday, 16 July 2024
- Lab
- Mistral AI
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Context tested | 256K tokens in-context retrieval | company |
No successor Mamba model appears in Mistral's 2024-2026 news feed.
Sources
This record was checked against its sources on 6 October 2026. How we check