News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
fnDevstral

Devstral

A 24B agentic coding model built by Mistral AI and All Hands AI

Devstral is an agentic large language model for software engineering, built by Mistral AI together with All Hands AI and released in 2025 as a research preview. Fine-tuned from Mistral Small 3.1, the 24 billion parameter model is trained to resolve real GitHub issues and runs inside code agent scaffolds such as OpenHands.

Share on WhatsApp

  • 24 billion parameters, fine-tuned from Mistral Small 3.1
  • 46.8% on SWE-Bench Verified with the OpenHands scaffold
  • Runs on a single RTX 4090 or a Mac with 32GB RAM
  • 128k token context window, Tekken tokenizer with 131k vocabulary
  • Weights released under the Apache 2.0 licence

What it is

Devstral is an agentic language model aimed at real software engineering work rather than isolated code completion. Mistral AI built it with All Hands AI and put it out in 2025 as a research preview. It has 24 billion parameters and is fine-tuned from Mistral Small 3.1, with the vision encoder stripped out so the model is text-only. Its job is to work through whole repositories: reading code spread across many files, spotting links between components and finding subtle bugs. It is trained to solve genuine GitHub issues and is meant to be driven by a code agent scaffold such as OpenHands or SWE-Agent, which supplies the interface to test cases.

How it came about

Mistral AI announced Devstral as a joint effort with All Hands AI, the company behind the OpenHands scaffold. The release was framed as a research preview, with the team saying it was already working on a larger agentic coding model for the following weeks. Enterprise customers wanting fine-tuning on private codebases, continued pre-training or distillation of the model's abilities into other systems were directed to Mistral AI's applied AI team. The company also said commercial models with extra capabilities, such as longer context or domain knowledge, would follow beyond what it gives to the community.

Inside the model

Devstral is fine-tuned from the Mistral Small 3.1 base model, so it inherits a context window of up to 128k tokens. Because it is used as a coding agent, the vision encoder was taken out before fine-tuning, leaving it text-only. It uses Mistral's Tekken tokenizer with a 131k vocabulary. The training target is real GitHub issues, and the model is meant to run inside agent scaffolds that handle the loop between the model and test cases. At 24 billion parameters it is small enough for a single RTX 4090 or a Mac with 32GB of RAM.

What it can do

On SWE-Bench Verified, a set of 500 real GitHub issues checked by hand, Devstral scores 46.8% using the OpenHands scaffold. Mistral AI says that beats the previous best open-source results by more than six percentage points, and that under the same scaffold it exceeds much larger models including Deepseek-V3-0324 (671B) and Qwen3 232B-A22B. Measured against models using any scaffold, it is reported as more than 20% ahead of GPT-4.1-mini. The model card lists 46.8 for Devstral, 40.6 for Claude 3.5 Haiku, 40.2 for SWE-smith-LM 32B and 23.6 for GPT-4.1-mini.

How to use it

Weights are downloadable under Apache 2.0 from Hugging Face, Ollama, Kaggle, Unsloth and LM Studio, and the licence allows commercial and non-commercial use and modification. The model is also served through Mistral AI's API as devstral-small-2505, priced the same as Mistral Small 3.1 at $0.1 per million input tokens and $0.3 per million output tokens. Local deployment is supported through vLLM, mistral-inference, transformers, llama.cpp and others, with OpenHands recommended as the agent scaffold.

Reception and debate

Mistral AI presents Devstral as the leading open-weight model on SWE-Bench Verified and reports that it outperforms several closed-source alternatives on that benchmark. The company describes the release as a research preview and asks users for feedback.

Why it matters

Devstral targets a gap the company identifies in coding models: handling whole codebases and real bug reports rather than single functions. Released openly under Apache 2.0, it is offered as a base for others to build on, customise and speed up autonomous software development. Mistral AI positions it for local use, privacy-sensitive enterprise repositories and as a model option inside coding IDEs, plugins and agent environments.

Questions people ask about Devstral

What is Devstral for?

It is built for software engineering tasks that span a whole repository, such as tracing relationships between files and finding subtle bugs. It is trained on real GitHub issues and is designed to be driven by a code agent scaffold like OpenHands or SWE-Agent rather than used for simple code completion.

Who made it?

Mistral AI built it in collaboration with All Hands AI, the company behind the OpenHands coding agent. It was released in 2025 as a research preview, with a larger agentic coding model said to be coming later.

How big is it and can I run it locally?

It has 24 billion parameters and a 128k token context window. Mistral AI says it is light enough for a single RTX 4090 or a Mac with 32GB of RAM, and weights can be downloaded for local deployment.

How do I get it?

Weights are free under Apache 2.0 on Hugging Face, Ollama, Kaggle, Unsloth and LM Studio. It is also on Mistral AI's API as devstral-small-2505, at $0.1 per million input tokens and $0.3 per million output tokens.

How well does it perform?

Mistral AI reports 46.8% on SWE-Bench Verified using the OpenHands scaffold, which it says is more than six points above prior open-source results. Under the same scaffold it is reported to beat larger models such as Deepseek-V3-0324 and Qwen3 232B-A22B.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Devstral yet, so the facts on this page come from the sources listed above: mistral.ai/news/devstral, huggingface.co/mistralai/Devstral-Small-2505. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.