Mistral releases Codestral 25.08 for enterprise coding
Mistral AI has released Codestral 25.08, saying it raises accepted completions by 30%, and is offering it with its embedding and agent models as one coding stack that enterprises can host themselves.
Mistral AI announced Codestral 25.08, the newest version of its code-completion model family. The company is presenting it as the base layer of a coding stack that enterprises can deploy in the cloud, in a VPC or on-premises.
Mistral reports these gains for Codestral 25.08 over earlier versions, measured in live IDE use on production codebases:
- 30% more accepted completions
- 10% more suggested code kept after acceptance
- 50% fewer runaway generations
- Better academic benchmark scores for short- and long-context fill-in-the-middle completion
In chat mode, Mistral says instruction following rose 5% on IF eval v8 and average MultiplE scores rose 5%.
The other layers of the stack are already available. Codestral Embed is a code-specific embedding model. Mistral says it beats OpenAI and Cohere embedding models on real-world code retrieval benchmarks and supports smaller outputs such as 256-dimension and INT8. Devstral is the agentic coding layer and runs on the OpenHands scaffold. Devstral Small 1.1 is an open-weight 24B model under Apache 2.0, and Mistral says it scores 53.6% on SWE-Bench Verified and runs on a single RTX 4090 or a Mac with 32 GB of RAM. Devstral Medium scores 61.6% and is offered through Mistral's API and enterprise partnerships.
Mistral argues that enterprises have been slow to adopt AI coding tools because most are SaaS-only, can't be customized and are hard to monitor. Its headline claim that customers cut development, review and testing time by 50% comes without named customers or a method. The excerpt also gives no pricing or availability details for Codestral 25.08.