Devstral 2

Devstral 2 is a 123-billion-parameter dense transformer model fine-tuned as an FP8 Instruct checkpoint for agentic software engineering workflows. Released by Mistral AI under a modified MIT license with openly available weights, it targets developers who need a model that can explore codebases, edit multiple files coherently, and reason about architecture-level dependencies. The instruct-style fine-tuning emphasizes chat, instruction following, and tool-driven task execution rather than open-ended generation, positioning it as a practical engine for coding assistants and autonomous SWE agents.

The model operates within a 256K-token context window, giving it room to hold sizable repositories in working memory while tracking frameworks, detecting failures, and retrying with corrections across complex edits. On software engineering benchmarks it posts 72.2% on SWE-bench Verified and 61.3% on SWE-bench Multilingual, placing it ahead of smaller open coder models and competitive with much larger frontier systems on real-world bug fixing. Marketed as a clear step-up from its predecessors with broader generalization across diverse prompts and coding environments, Devstral 2 suits teams building agentic coding tools, automated refactoring pipelines, and enterprise-scale code modernization workflows that benefit from open-weight deployment.