Mistral Releases Codestral 25.08 with 30% Higher Completion Acceptance, Ships Full Enterprise Coding Stack
Mistral AI released Codestral 25.08, showing 30% more accepted code completions and 10% higher retention rates. The company also shipped Devstral Small, a 24B-parameter agentic coding model scoring 53.6% on SWE-Bench Verified, alongside new embedding and IDE integration tools aimed at enterprise deployment.
Mistral Releases Codestral 25.08 with 30% Higher Completion Acceptance, Ships Full Enterprise Coding Stack
Mistral AI released Codestral 25.08, the latest version of its code generation model, alongside a complete coding stack targeting enterprise deployments. The update shows measurable improvements in code completion quality, according to the company.
Codestral 25.08 Performance Metrics
Mistral claims Codestral 25.08 delivers:
- 30% increase in accepted completions compared to prior versions
- 10% more code retained after suggestion
- 50% fewer runaway generations
- 5% improvement on instruction-following benchmark IFEval v8
- 5% gain in code abilities on MultiplE benchmark
The model is optimized for fill-in-the-middle (FIM) completion and supports deployment across cloud, VPC, or on-premises environments.
Devstral: Agentic Coding Models
Mistral introduced Devstral, a family of models built for multi-file coding tasks. Devstral Small (24B parameters, Apache-2.0 license) scores 53.6% on SWE-Bench Verified, while Devstral Medium reaches 61.6%. According to Mistral, these scores exceed Claude 3.5, GPT-4.1-mini, and other open models on the benchmark.
Devstral Small runs on a single Nvidia RTX 4090 or Mac with 32GB RAM. The model is open-weight, allowing enterprises to fine-tune on proprietary codebases or integrate into CI/CD pipelines without licensing restrictions. Devstral Medium is available through Mistral's API and enterprise partnerships.
Codestral Embed for Code Search
The stack includes Codestral Embed, a code-specific embedding model. Mistral claims it outperforms OpenAI and Cohere embeddings on code retrieval benchmarks. The model offers configurable dimensions (including 256-dim INT8) and supports private deployment for monorepo and poly-repo search.
IDE Integration and Enterprise Controls
All capabilities are accessible through Mistral Code, a plugin for JetBrains and VS Code. Features include:
- Inline completions via Codestral 25.08
- One-click automations powered by Devstral
- Semantic search backed by Codestral Embed
- Context awareness from Git diffs and terminal history
The plugin supports cloud, VPC, and on-premises deployment (general availability Q3 2025). Mistral offers SSO, audit logging, and usage controls through the Mistral Console. No telemetry is mandatory, and inference can run without external API calls.
Enterprise Deployment Options
Mistral's approach addresses several enterprise requirements:
- VPC, on-premises, and air-gapped deployment options
- Model weights access for customization
- Unified observability and audit trails
- Integration with internal CI/CD and knowledge bases
Pricing for Codestral 25.08 and Devstral models was not disclosed.
What This Means
Mistral is positioning itself as an enterprise-first alternative to SaaS-only coding assistants. The combination of open-weight models (Devstral Small), private deployment options, and integrated tooling addresses compliance and customization requirements common in regulated industries. The 30% improvement in completion acceptance, if validated independently, represents meaningful progress in code suggestion quality. However, real-world enterprise adoption will depend on whether the stack's complexity and integration requirements match organizations' deployment capabilities. SWE-Bench Verified scores above 50% suggest Devstral is competitive with frontier models on multi-file coding tasks, though these benchmarks don't capture all aspects of production coding workflows.
Related Articles
Hugging Face Transformers v5.16.1 Adds Support for GLM-5.3-Flash, a 320B-Parameter Multimodal MoE Model
Hugging Face's transformers v5.16.1 release adds support for GLM-5.3-Flash, a 320B total-parameter (18B active) multimodal mixture-of-experts model. Zhipu AI claims it outperforms GLM-5.2 while approaching Claude Opus 4.8 on coding and agentic benchmarks at one-tenth the cost.
Cline Desktop v0.0.19 Fixes Memory Leak That Ballooned Process to Tens of Gigabytes
Cline Desktop v0.0.19 fixes a memory leak where session status updates carried full conversation transcripts to every connected client, ballooning process memory to tens of gigabytes on long tasks. The release also adds seven new model providers and changes default models for several existing ones.
Ollama v0.33.0 Fixes KV Cache Bug That Forced Full Reprocessing of 46K-Token Prompts
Ollama's v0.33.0 pre-release adds direct model management inside Claude Desktop's menu bar and fixes a caching bug that could force reprocessing of tens of thousands of tokens. The release also disables a Claude Code system message that was breaking KV cache hits on every request.
OpenAI Reinstates 5-Hour Usage Limit for ChatGPT Plus Codex and Work Tiers
OpenAI will reinstate a five-hour usage limit for Codex and ChatGPT Work on Plus subscriptions starting August 25, 2026, after weeks of running only a weekly cap. Pro $100 and Pro $200 plans remain exempt from the change for the foreseeable future.
Comments
Loading...