Devstral
Last reviewed
Jun 8, 2026
Sources
5 citations
Review status
Source-backed
Revision
v1 · 1,458 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 8, 2026
Sources
5 citations
Review status
Source-backed
Revision
v1 · 1,458 words
Add missing citations, update stale details, or suggest a clearer explanation.
Devstral is a family of open-weight and API large language models specialized for agentic software engineering, developed by Mistral AI in collaboration with All Hands AI, the team behind the OpenHands coding-agent platform (formerly OpenDevin). Unlike conventional code generation models that produce isolated completions or snippets, Devstral is tuned to operate inside an agent scaffold: it explores a codebase, reasons about how components relate, uses tools, and makes multi-file edits in order to resolve real GitHub-issue-style tasks.[1][2] The first model, Devstral Small, was released on May 21, 2025 under the permissive Apache 2.0 license, followed in July 2025 by an updated Devstral Small 1.1 and a stronger commercial model, Devstral Medium.[1][3] Mistral reported that Devstral models topped open models on SWE-bench Verified, the standard benchmark for autonomous software-engineering agents.[1][3]
Devstral targets the task of agentic software engineering, in which a model acts as an autonomous coding agent rather than an autocomplete assistant. Given a natural-language description of a bug or feature request and access to a repository, the agent must locate the relevant files, understand context across a large codebase, modify code, and arrive at a solution that passes hidden test cases.[1] To do this, the model does not run in isolation; it runs over a code-agent scaffold such as OpenHands or SWE-agent, which defines the interface between the model and the environment, including how the model reads files, runs shell commands, and applies edits.[1]
The flagship open model, Devstral Small, has 24 billion parameters, making it compact enough to run on a single high-end consumer GPU such as an NVIDIA RTX 4090, or on a Mac with 32 GB of unified memory.[1] This local-deployment property, combined with an Apache 2.0 license, distinguished Devstral from the closed agentic-coding systems offered by competitors and made it attractive for developers who want to run a capable software-engineering agent on their own hardware.[1][4]
Devstral was built jointly by Mistral AI, a Paris-based AI company, and All Hands AI, the organization that develops OpenHands.[1] OpenHands began life as OpenDevin, an open-source reimplementation of the autonomous-developer concept popularized by the Devin agent, and provides scaffolds that define how a model interacts with a repository and its test suite.[1][2] All Hands AI's experience with these scaffolds shaped the way Devstral was trained: the model was optimized to perform well when driven by the OpenHands agent loop, where it issues tool calls, inspects results, and iterates toward a fix.[1]
Devstral is distinct from Codestral, Mistral's earlier code model. Codestral is oriented toward code completion and fill-in-the-middle generation across many programming languages and is licensed under Mistral's non-production license. Devstral, by contrast, is purpose-built as a software-engineering agent that executes multi-step tasks, and the open Small variant is released under Apache 2.0.[1][4] The two models reflect a division of labor in Mistral's lineup: Codestral for fast in-editor completion, Devstral for autonomous issue resolution.
Devstral Small is a fine-tune of Mistral's general-purpose Mistral Small 3.1 family, specifically the Mistral-Small-3.1-24B-Base-2503 base model. During fine-tuning the vision encoder of the multimodal base was removed, so Devstral Small is a text-only model focused on code.[5]
Mistral has released the Devstral family in two waves. The initial open release in May 2025 introduced Devstral Small (internally Devstral-Small-2505). The July 2025 update introduced Devstral Small 1.1 (Devstral-Small-2507), an improved open model, alongside Devstral Medium (devstral-medium-2507), a larger commercial model offered through Mistral's API and enterprise products.[1][3]
| Property | Devstral Small (2505) | Devstral Small 1.1 (2507) | Devstral Medium (2507) |
|---|---|---|---|
| Released | May 21, 2025 | July 10, 2025 | July 10, 2025 |
| Parameters | 24 billion | 24 billion | Not disclosed |
| Base model | Mistral Small 3.1 (24B) | Mistral Small 3.1 (24B) | Not disclosed |
| Context window | 128,000 tokens | 128,000 tokens | Not disclosed |
| Tokenizer | Tekken (131k vocabulary) | Tekken (131k vocabulary) | Not disclosed |
| License | Apache 2.0 (open weights) | Apache 2.0 (open weights) | Commercial / API only |
| SWE-bench Verified (Mistral, OpenHands scaffold) | 46.8% | 53.6% | 61.6% |
| Availability | Hugging Face, API, local | Hugging Face, API, local | API, on-premise, fine-tuning |
Sources: Mistral AI announcements and Hugging Face model card.[1][3][5]
Devstral Small uses the Tekken tokenizer with a vocabulary of roughly 131,000 tokens and supports a 128,000-token context window, allowing the agent to hold large amounts of code in context while working.[5] Mistral describes Devstral Small 1.1 as offering improved performance and better generalization across different prompts and coding environments compared with the original release, while keeping the same 24B architecture. Version 1.1 also added explicit support for both Mistral function-calling and XML tool-call formats, improving its versatility across agentic scaffolds beyond OpenHands.[3]
Devstral Medium is positioned as a higher-capability model available only through Mistral's commercial channels, including the Mistral API, on-premise deployment, custom fine-tuning via Mistral's fine-tuning API, and the Mistral Code product for enterprises. Mistral did not publish a parameter count for Devstral Medium.[3]
Mistral evaluated Devstral primarily on SWE-bench Verified, a human-validated subset of SWE-bench consisting of real GitHub issues paired with tests that verify whether a proposed patch resolves the problem. All of Mistral's headline scores were obtained using the OpenHands scaffold provided by All Hands AI, and should be read as the company's own reported results rather than independent measurements.[1][3]
According to Mistral:
These comparisons depend heavily on the agent scaffold and evaluation harness, and benchmark numbers for agentic coding can vary between setups, so the figures are best understood as Mistral's claims under a specific OpenHands configuration.[1][3]
Devstral Small is distributed as an open-weight model under the Apache 2.0 license, which permits free use for both research and commercial purposes. The weights are hosted on Hugging Face (as Devstral-Small-2505 and the later Devstral-Small-2507) and are downloadable for local use.[5] The model can be run with common open-source inference stacks, and Mistral and the community have made it available through frameworks and tools including vLLM, mistral-inference, Hugging Face Transformers, Ollama, LM Studio, llama.cpp, Kaggle, and Unsloth.[5]
Both Devstral Small and Devstral Medium are also offered through Mistral's API. At launch Mistral priced the Small model at $0.10 per million input tokens and $0.30 per million output tokens, and Devstral Medium at $0.40 per million input tokens and $2.00 per million output tokens.[1][3] Devstral Medium is additionally available for on-premise deployment, custom fine-tuning, and through the Mistral Code enterprise offering, but its weights are not released openly.[3]
Devstral was significant as a leading open agentic coding model. At a time when the strongest software-engineering agents were closed systems built on models such as Claude and GPT, Devstral demonstrated that a relatively small 24B open-weight model, paired with a capable scaffold like OpenHands, could lead open-model results on SWE-bench Verified while remaining light enough to run on a single consumer GPU or a laptop.[1][4] This combination of openness, modest hardware requirements, and a permissive Apache 2.0 license made it a practical foundation for developers and companies building their own coding agents without depending on a proprietary API.[4]
The release also reflected a broader trend toward specialization within the Mistral lineup and the wider field: rather than relying on a single general model, vendors increasingly shipped distinct models for code completion (Codestral) and for autonomous, tool-using software engineering (Devstral). By collaborating with All Hands AI, Mistral tied its model directly to an open agent platform, reinforcing the idea that agentic coding performance is a property of the model and scaffold together rather than the model alone.[1][2]