Mistral Releases Devstral, an Apache 2.0 Agentic Coding Model
Mistral AI
- Models
- Agents
- Open source
- #devstral
- #swe-bench
- #apache 2.0
Mistral AI announced Devstral, an agentic LLM for software engineering built with All Hands AI. Mistral reports 46.8% on SWE-Bench Verified, more than 6 points above prior open-source state of the art, under the Apache 2.0 license and runnable on a single RTX 4090 or a 32GB Mac. It is offered via API as devstral-small-2505 and positioned as a research preview.
What Mistral announced
Mistral AI introduced Devstral, an agentic large language model aimed at software engineering, developed in collaboration with All Hands AI. The distinction Mistral draws is between models that handle atomic coding tasks such as code completion and models that can work through real-world issues: locating code inside a larger repository, tracing relationships between components, and finding subtle bugs. Devstral is trained to solve real GitHub issues and runs over agent scaffolds such as OpenHands or SWE-Agent, which define the interface between model and test cases. Mistral reports 46.8% on SWE-Bench Verified, a screened set of 500 real GitHub issues, saying this beats prior open-source state of the art by more than 6 points. Under the OpenHands scaffold, Mistral says Devstral exceeds much larger models including Deepseek-V3-0324 and Qwen3 232B-A22B, and surpasses GPT-4.1-mini by over 20%. The model is released under Apache 2.0, is described as light enough for a single RTX 4090 or a 32GB Mac, and is offered on Mistral's API as devstral-small-2505.
A scenario for a platform engineering lead
Consider a platform lead at an organization with repositories that cannot leave its network. Because Mistral describes Devstral as small enough for local hardware and licensed for customization, such a team could try pairing it with a self-hosted OpenHands-style scaffold on an internal workstation or server, then point the agent only at non-production branches. The premise to test is Mistral's claim about resolving real issues in large codebases, not a guarantee of outcome. This scenario stays within what the source supports: local deployment, an agent scaffold, and an Apache 2.0 release that permits building on and customizing the model.
How to run a bounded trial
A team could select a small backlog of already-closed, well-specified issues with known merged fixes and run the agent in a sandbox, recording baseline human time per issue before starting. Success would be measured against that baseline: how often the agent's patch passes existing tests, how many issues need human correction, and review time per patch. Failure checks should include patches that pass tests but change behavior outside the issue scope, and any signs the scaffold rather than the model drives results, since Mistral notes results depend on the scaffold. Keep a human reviewer as the merge gate throughout, and stop the trial if failure modes are not understood.
Trade-offs and when to hold back
Local self-hosting avoids sending code off-site but requires hardware and scaffold maintenance; the API route trades that for per-token costs and data leaving the network. Mistral calls Devstral a research preview and says a larger agentic coding model is coming, so a team may reasonably wait for that or re-evaluate later. The benchmark figures come from the publisher and were not independently tested here, and performance is scaffold-dependent, so comparisons with other models under custom scaffolds are not apples-to-apples. Teams without a sandboxed test process or human review capacity should not adopt agentic patching yet.
Context and limitations
- Based on a short publisher excerpt; benchmark numbers are Mistral's own and were not independently verified here.
- Mistral labels Devstral a research preview and says a larger agentic coding model is planned.
- Performance depends on the agent scaffold; results cited use OpenHands or any-scaffold comparisons.
- Based on bounded source text, not independent testing. Practical suggestions are editorial proposals.