📢 Update: Mistral Medium 3.5 has replaced Devstral 2 as the default model in Vibe CLI. See the Medium 3.5 complete guide and Vibe 2.0 remote agents guide.

Devstral 2 is Mistral AI’s dedicated coding agent model — a 123B parameter model that scores 72.2% on SWE-bench Verified, matching Claude Opus 4.6. It’s open-weight under a modified MIT license and designed for autonomous coding tasks.

Key facts

• 123B dense parameters — runs on a single server node
• 256K context window — the largest among coding models
• 72.2% SWE-bench — matches Claude Opus, beats GPT-5.4
• Modified MIT license — open for commercial use
• Also available as Devstral Small (24B) — runs on consumer hardware

What Devstral 2 does

Devstral 2 is built for agentic coding — tasks where the AI needs to autonomously plan, execute, and iterate. This includes:

• Bug fixing — reads error logs, traces the issue, applies a fix
• Feature implementation — takes a spec and builds it across multiple files
• Refactoring — restructures code while maintaining behavior
• Test generation — writes comprehensive test suites for existing code
• Code review — analyzes PRs and suggests improvements

Unlike Codestral (which is optimized for fast autocomplete), Devstral 2 is designed for complex multi-step tasks that require deep reasoning about code architecture.

Architecture

Devstral 2 is a dense transformer — all 123B parameters activate for every token. This is different from MoE models like Qwen 3.5 or DeepSeek V3 that only activate a subset. The dense architecture provides more consistent quality across different task types but requires more compute per token.

The 256K context window means it can process approximately 500-800 files of typical source code in a single pass, making it suitable for understanding entire microservice architectures or large monorepos.

Devstral 2 vs Codestral

Both are from Mistral but serve different purposes:

• Devstral 2 — for agent tasks (refactoring, bug fixing, building features)
• Codestral — for autocomplete (tab completions in your IDE)

The ideal setup is using both: Codestral for real-time inline suggestions as you type, and Devstral 2 for larger tasks you delegate to an AI agent.

How to use Devstral 2

• Vibe CLI — Mistral’s native terminal coding tool
• Aider — open-source CLI with Mistral support
• OpenCode — via Mistral API
• Mistral API — direct integration at $2/$6 per million tokens

Devstral Small (24B)

For developers who want to run locally, Devstral Small 2 is a 24B parameter version that fits on consumer hardware (16GB+ RAM). It sacrifices some quality compared to the full 123B model but still outperforms most open-source alternatives at its size.

ollama pull devstral-small:24b

FAQ

Can I run Devstral 2 locally?

The full 123B model requires significant hardware — approximately 80GB+ of VRAM across one or more GPUs. For local use, Devstral Small (24B) is the practical choice, running comfortably on machines with 16GB+ RAM via Ollama.

How does Devstral 2 compare to Claude Code?

Both score nearly identically on SWE-bench (72.2% vs 72.1%), so coding quality is comparable. The key difference is that Devstral 2 is a model you can access via API or self-host, while Claude Code is a complete CLI tool with built-in workflows. You can use Devstral 2 through Vibe CLI or Aider for a similar experience.

Is Devstral 2 truly open source?

Devstral 2 uses a modified MIT license that allows commercial use, modification, and redistribution. The weights are available for download. However, it’s “open-weight” rather than fully open-source — the training data and training code are not released, only the model weights.

Learn more

• Devstral 2 Complete Guide — full architecture and setup
• Devstral Small 2 Guide — the 24B version for local use
• Devstral 2 vs GLM-5.1 vs Codestral — comparison