Last updated 07-24-2026
Category:
Reviews:
Join thousands of AI enthusiasts in the World of AI!
Codestral Mamba - Mistral AI
Codestral Mamba is Mistral AI's open-weight code model built on the Mamba architecture rather than a standard Transformer. Released in July 2024, it targets code productivity with linear-time inference and strong performance on long contexts.
The 7.3B-parameter instructed model supports in-context retrieval tested up to 256k tokens, making it suited for local coding assistants and long codebase interactions. Weights are published on Hugging Face under Apache 2.0, and you can deploy via mistral-inference, TensorRT-LLM, or test through Mistral la Plateforme as codestral-mamba-2407.
Mamba's sequence modeling differs from Transformers by scaling inference time linearly with input length, which Mistral positions as an advantage for fast responses on large prompts during development workflows.
Mamba architecture delivers linear-time inference for long code prompts
7.3B instructed parameters tuned for code and reasoning tasks
Tested in-context retrieval up to 256k tokens for large codebases
Apache 2.0 open weights free to use, modify, and redistribute
Deploy locally via mistral-inference SDK, TensorRT-LLM, or llama.cpp support
Available on Mistral la Plateforme as codestral-mamba-2407 for API testing
Designed as a local code assistant with fast responses at any input length
Developed with input from Mamba architecture researchers Albert Gu and Tri Dao
Open Apache 2.0 weights enable self-hosting without commercial license fees.
Mamba linear inference scales better on very long code contexts than many Transformers.
Backed by Mistral AI with Hugging Face distribution and multiple deployment paths.
Announcement page is a blog post, not a dedicated product dashboard.
Smaller 7.3B size may trail largest frontier models on complex reasoning benchmarks.
What is Codestral Mamba?
Codestral Mamba is Mistral AI's open-weight code model based on the Mamba architecture. It has 7.3B parameters and is optimized for code generation and reasoning with long context windows.
Is Codestral Mamba free?
Yes. Model weights are released under the Apache 2.0 license for free use, modification, and distribution. API access on la Plateforme follows Mistral's platform pricing.
How do I run Codestral Mamba locally?
Download weights from Hugging Face and deploy using the mistral-inference SDK (based on the Mamba reference implementation), TensorRT-LLM, or supported local runners like llama.cpp.
What context length does Codestral Mamba support?
Mistral tested in-context retrieval capabilities up to 256k tokens, positioning the model for long code files and extended assistant sessions.
How is Codestral Mamba different from Codestral 22B?
Codestral Mamba uses the Mamba architecture with Apache 2.0 open weights. Codestral 22B is a larger Transformer-based model available under Mistral's commercial or community licenses.