Codestral 25.08
2025-07-30 · Publisher : Mistral AI · Domain : Artificial Intelligence
Mistral AI announces Codestral 25.08, an update to its code completion model (+30% accepted completions, 50% fewer runaway generations), as part of an enterprise coding stack.
What changes
- +30% accepted completions, +10% code retained after suggestion, 50% fewer "runaway" generations, measured in real-world use in IDEs.
- Improvement on academic FIM benchmarks in short and long context.
- Improved chat mode: +5% in instruction following (IF eval v8) and +5% in code (MultiPL-E average).
- Presented with Codestral Embed, Devstral and Mistral Code as an integrated stack deployable in the cloud, in a VPC or on-premises.
### Specifications - Context: 128k tokens (documentation); text and code, FIM endpoint. - License: proprietary (Premier model). - Availability: La Plateforme API; deployment in the cloud, in a VPC or on-premises. - API identifier: `codestral-2508`. ### Pricing | Model | Input (USD / M tokens) | Output (USD / M tokens) | |---|---|---| | Codestral 25.08 | 0.3 | 0.9 | Prices recorded from the pricing page archived on July 30, 2025 and from the documentation. ### Benchmarks Improvements published by the vendor (internal measurements in production, no absolute scores): +30% accepted completions, +10% code retained, 50% fewer runaway generations, +5% IF eval v8, +5% MultiPL-E. ### Key takeaways A code completion update validated in real-world use, especially interesting for companies that want a coding assistant deployable within their own infrastructure.
Links
Tags : LLM, Mistral, Codestral, Code, Complétion