Codestral Mamba
7B Mamba2 code model with a 256K context window under Apache 2.0.
About
Mistral released Codestral Mamba on July 16, 2024: a 7B code model built on the Mamba2 state-space architecture instead of transformers, developed with Mamba researchers Albert Gu and Tri Dao. Linear-time inference means fast responses even on very long inputs, and the context window is 256K tokens — handy for whole-repo questions. Team-reported: it beat CodeLlama 7B, CodeGemma, and DeepSeek's coder on HumanEval. Apache 2.0, so genuinely free for commercial use, and it runs on modest local hardware. Honest caveats: Mamba coders are less battle-tested in the assistant ecosystem than transformer coders, and the HumanEval comparisons are Mistral's own.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Team-reported: outperformed CodeLlama 7B, CodeGemma-1.1 7B, and DeepSeek Coder on HumanEval.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (verified tier, metrics-v1). Evidence: July 16, 2024 release verified via press; Apache 2.0 verified.; 7B Mamba2 code model, 256K context verified via Mistral announcements and OpenRouter.; Team-reported HumanEval wins over CodeLlama 7B / CodeGemma labeled as Mistral's own.; Ecosystem maturity caveat included..