Codestral Mamba 7B
Codestral Mamba uses Mamba 2 instead of the Transformer architecture of the original Codestral. Mistral released its instruction checkpoint for local coding and architecture research.
A different sequence architecture
The release was designed with help from Albert Gu and Tri Dao. Mistral tested in-context retrieval to 256K tokens. That evaluation is a useful bound: the architecture’s theoretical ability to process longer sequences does not establish unlimited reliable context.Codestral Mamba 7B release announcement
Local inference requirements
The official card uses Mistral’s inference library with mamba-ssm and causal-conv1d, or the original Mamba implementation. This release needs a state-space implementation; a server that supports only Transformer checkpoints is insufficient. The checkpoint name is Mamba-Codestral-7B-v0.1.Codestral Mamba 7B official model card
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | |
| Access | Downloadable |
| Model identifiers | Mamba-Codestral-7B-v0.1open-codestral-mambacodestral-mamba-2407 |
| Context window | 256,000 tokens The model catalog advertises 256k context for this version. This is not a measured recall limit. Codestral Mamba 7B model and lifecycle record |
| Weight license | Apache 2.0Codestral Mamba 7B official model card |
| Access and lifecycle | Hosted API retired on 6 June 2025. Downloadable weights remain listed.Codestral Mamba 7B model and lifecycle recordCodestral Mamba 7B official model card |
| Parameters | 7,285,403,648 in the release announcementCodestral Mamba 7B release announcement |