Devstral
By Mistral AI (with All Hands AI)
Devstral is an open-weight large language model from Mistral AI, developed in collaboration with All Hands AI, specifically optimized for agentic software engineering tasks such as autonomously navigating codebases, applying multi-file…
Definition
Devstral is an open-weight large language model from Mistral AI, developed in collaboration with All Hands AI, specifically optimized for agentic software engineering tasks such as autonomously navigating codebases, applying multi-file edits, and resolving GitHub issues.
Overview
Devstral, released in mid-2025, targets a distinct niche from general-purpose code-completion models like Codestral: rather than optimizing primarily for inline code suggestions, it's tuned specifically for agentic coding workflows, where a model operates somewhat autonomously within a coding agent framework — reading and exploring a codebase, deciding which files are relevant, planning multi-step changes, executing edits across multiple files, and iterating based on test or build feedback. Mistral developed Devstral in collaboration with All Hands AI, the team behind the OpenHands (formerly OpenDevin) open-source coding agent framework, and the model's training and evaluation were closely tied to benchmarks like SWE-bench, which measures a model's ability to resolve real-world GitHub issues by generating correct patches against actual open-source repositories. Devstral was released as an open-weight model (under the Apache 2.0 license for the initial Devstral Small variant), making it usable within self-hosted agentic coding tools rather than only via a proprietary API. Devstral reflects a broader 2024-2025 industry shift toward agentic coding as a distinct model specialization, alongside offerings like Anthropic's Claude models tuned for agentic tool use, OpenAI's Codex-related models, and other open agentic coding models. Its differentiation is being specifically benchmarked and tuned for the multi-step, tool-using, codebase-navigating workflow that autonomous coding agents require, rather than the simpler single-turn completion or generation task that earlier code models like Codestral primarily targeted.
Key Concepts
- Optimized specifically for agentic software engineering workflows, not just code completion
- Developed in collaboration with All Hands AI (OpenHands coding agent framework)
- Evaluated on SWE-bench, a benchmark for resolving real GitHub issues with correct patches
- Released as an open-weight model under Apache 2.0 (Devstral Small)
- Designed to navigate multi-file codebases and plan multi-step changes autonomously
- Integrable with open-source and third-party agentic coding tools