Overview
Devstral 2 is the second generation of Mistral AI's open-weight coding model family, announced on December 9, 2025. Mistral AI is a French startup founded in 2023 by former DeepMind and Meta researchers and focused on open and commercial large language models. Devstral 2 ships in two sizes: the flagship Devstral 2 (123B parameters) and Devstral Small 2 (24B parameters). Both are dense transformers built for software-engineering agents that autonomously explore codebases, edit multiple files, and resolve issues. They solve the problem of giving developers a capable, license-permissive coding model that can run either through Mistral's API or locally on modest hardware, lowering the barrier for production-grade AI coding assistance.
Key Features
- 123B (Devstral 2) and 24B (Devstral Small 2) dense transformers
- 256K context window on both sizes
- 72.2% on SWE-bench Verified (Devstral 2); 68.0% (Small 2)
- Up to 7x more cost-efficient than Claude Sonnet on real-world tasks (per Mistral)
- Mistral Vibe CLI: open-source, terminal-native coding agent
- Devstral Small 2 supports image inputs for multimodal agents
- On-prem deployment and custom fine-tuning supported
Pros (based on real user feedback)
- New open-source state of the art for code agents at a fraction of competitor model size
- Strong cost efficiency, and free during the introductory API period
- Small 2 runs locally on consumer GPUs for private, on-device use
- Praised by Cline and Kilo Code for smooth, reliable tool-calling
- Permissive licenses enable commercial use and fine-tuning
Cons (based on real user feedback)
- In independent human evaluations, Claude Sonnet 4.5 was still significantly preferred, showing a remaining gap with top closed models
- The 123B flagship requires at least 4 H100-class GPUs to deploy, a high hardware bar
- The "free API" window is promotional and pricing will rise afterward
- The small model, while capable, trails the flagship on the hardest tasks
Who It's For
Devstral 2 is aimed at engineering teams and developers who want a strong, self-hostable coding agent without the per-token costs of closed APIs. The 123B model fits organizations with GPU clusters; the 24B model is ideal for hobbyists, startups, and privacy-conscious users who want local, on-device code assistance. It is less suited to users who need guaranteed parity with the very best closed coding models.
Verdict
Devstral 2 is a standout open-weight coding release: compact, fast, and license-friendly, with a free API window that makes evaluation painless. The 24B variant in particular is a compelling local-coding option. Just plan around the eventual paid API tier and the fact that the largest closed models still edge it out on quality.