Devstral

Devstral — explained in more detail

Mistral AI introduced Devstral on May 21, 2025 as an agentic model for software engineering, developed in collaboration with All Hands AI. Unlike pure code completers (such as Codestral), Devstral is trained to solve whole tasks within a code-agent scaffold — for example working on real GitHub issues, navigating across multiple files and calling tools. It runs over agent scaffolds such as OpenHands or SWE-Agent. The first version had 24 billion parameters, was released as open weights under the Apache 2.0 license, and reached 46.8% on SWE-bench Verified — the best score among open models at the time.

At the end of 2025 came Devstral 2 (December 9, 2025) in two sizes: a 123B flagship and Devstral Small 2 with 24B parameters, scoring 72.2% on SWE-bench Verified.

Example / Practical use

Devstral is deliberately lean enough to run locally — the smaller variants work on a single RTX 4090 or a Mac with 32 GB of RAM. That makes it attractive to developers and teams with data-sovereignty requirements who want to run a coding agent without a cloud API. The typical use: automated bug fixes and feature tasks directly against a repository, driven by an agent framework. The Apache 2.0 and other permissive licenses also allow commercial use.

Distinction

Devstral differs from Codestral in its task: Codestral completes code in the editor, while Devstral solves tasks agentically across an entire project. Among open models, Devstral positions itself as a specialized SWE-agent model — not a generalist but a locally runnable coding agent with strong SWE-bench scores.

More on this topic

See everything in one place:Mistral→

GlossaryMistral 7BMistral 7B (September 2023) is Mistral AIs first open-weight model — a 7.3B-parameter model under Apache 2.0 that at the time outperformed the larger Llama 2 13B on benchmarks.GlossaryMistral Medium 3Mistral Medium 3 is a proprietary AI model from Mistral AI released in May 2025. As an enterprise model built for high performance at low operating cost, it offers around 128,000 tokens of context and can also be run on-premise or in your own VPC.GlossaryMixtral 8x7BMixtral 8x7B (December 2023) is Mistral AIs open Sparse-MoE model with 8 experts — 46.7B parameters, 12.9B active per token, 32K context, Apache 2.0 license.