Codestral Mamba

// Overview

Codestral Mamba is an open 7-billion-parameter code model from Mistral AI, released on July 16, 2024, alongside the math model Mathstral as part of Mistral's specialist open-model line. Unlike Mistral's Transformer-based Codestral 22B, Codestral Mamba is built on the Mamba2 architecture, a state-space model that offers linear-time inference and the theoretical ability to handle sequences of unbounded length. Mistral tested it for in-context retrieval up to 256k tokens, making it well suited to running through long files and large codebases without the quadratic slowdown a Transformer hits as the prompt grows.

The model has exactly 7,285,403,648 parameters and was released under the permissive Apache 2.0 license, free to use, modify, and self-host. Mistral published it both on its la Plateforme API under the name open-codestral-mamba (also testable as codestral-mamba-2407) and as open weights on Hugging Face at mistralai/Mamba-Codestral-7B-v0.1. It can be deployed with Mistral's mistral-inference SDK or NVIDIA's TensorRT-LLM, and it was positioned as a lightweight local code assistant alternative to the larger Codestral 22B.

Mistral retired the open-codestral-mamba endpoint from la Plateforme on June 6, 2025, pointing API users to the newer Codestral model. The open weights, however, remain available under Apache 2.0, so the model can still be downloaded, fine-tuned, and self-hosted indefinitely. It is best understood today as a historical, research-oriented release that demonstrated Mamba-style architectures could match Transformer code models on standard benchmarks at the 7B scale.

// Benchmarks

1. HumanEval (Python, pass@1)75%
2. MBPP68.5%
3. Spider (text-to-SQL)58.8%
4. CruxEval57.8%
5. HumanEval C++59.8%
6. HumanEval Java57%
7. HumanEval JavaScript61.5%
8. HumanEval Bash31.1%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• Mamba2 architecture gives linear-time inference, so throughput stays steady even as the input grows long
• Tested for in-context retrieval up to 256k tokens — useful for reasoning over large files and codebases
• Strong code benchmarks for a 7B model (75.0% HumanEval Python), competitive with much larger Transformer code models
• Apache 2.0 open weights: free to self-host, fine-tune, and redistribute with no usage restrictions
• Small enough (~7.3B params) to run locally as a code-completion assistant

// Best for

• Local, low-latency code autocompletion and inline coding assistance
• Generating and explaining code where a long context (large files, many imports) must stay in the prompt
• Research and experimentation with non-Transformer (state-space) architectures for code
• Self-hosted or air-gapped coding tools that need a permissively licensed model
• Fine-tuning a compact open code model for a specific language or codebase

// How to access

// Specialist open models (Mathstral / Codestral Mamba) — every version

The full lineage of the Specialist open models (Mathstral / Codestral Mamba) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.