The model is aimed at coders. It expands the coding context window and focuses on enhancing current code. The LLM also shows what is possible with agentic technology.
French AI startup Mistral AI on Wednesday introduced a new agentic LLM for software engineering tasks.
The startup launched Devstral in partnership with All Hands AI, an agentic AI startup.
Mistral said the new model is designed to solve "real GitHub issues" and runs over code agent scaffolds like OpenHands and SWE-agent.
A scaffold is a temporary structure or workflow that enables developers to achieve a particular goal. The SWE-agent system powers large language models (LLMs) acting as software engineering agents to perform engineering tasks.
The LLM also tackles the problem of contextualizing code within a large codebase and identifying different components and bugs.
Larger context window
What Mistral essentially did with Devstral was enlarge the context window so the LLM can "consider more of the interdependencies within the current codebase when creating new code," said Torsten Volk, an analyst at Omdia, a division of Informa TechTarget.
"This is similar to human coders, who need to make sure they use consistent authentication logic, error handling, naming conventions, data validation, logging patterns or API structures," Volk said.
He added that the model must still perform its job to be optimal.
"The model still has to consistently identify the important part within these increasingly large context windows, which is a bigger challenge than just increasing the window size," he continued.
Another interesting part of what Mistral is doing with Devstral is a feature that enables the model to focus on enhancing existing code without having to rewrite entire files.
"This allows developers to enhance current code without risking losing existing features in the process, which has proven to be a significant problem in the past," Volk said.
Nature of agentic AI and challenges
Besides being helpful to developers, the model speaks to a larger trend in the generative AI market.
Devstral's completely agentic nature speaks to how agentic AI has shifted the narrative about what can be accomplished with generative AI, said Bradley Shimmin, an analyst at The Futurum Group.
"That ability for these models to take on a higher-level task, higher-level reasoning tasks, is the future," Shimmin said.
Being viewed as enterprise-grade and enterprise-ready is their biggest challenge.
Bradley ShimminAnalyst, Futurum Group
He added that while Mistral -- which competes against other independent generative AI vendors such as OpenAI, Anthropic and Cohere as well as the tech giants -- is aiming to drive innovation for the open source community, it still has to work on its positioning.
"Being viewed as enterprise-grade and enterprise-ready is their biggest challenge," Shimmin continued. "Where do they fit?"
He added that it is unclear whether the vendor aims to replace copilots from OpenAI and Microsoft or take a largely consumer angle like Gemini but with enterprise recognition.
"It's kind of neither of them," he continued. "It's its own space. The biggest challenge is how they evolve over time to strike that balance between consumer and enterprise focus and capability."
Mistral said Devstral is light enough to run on a single Nvidia RTX 4090 GPU or a Mac with 32 GB RAM.
It is available for free under the Apache 2.0 license.
Meanwhile, OpenAI also added new built-in tools to its Responses API, an API for building agentic applications. The tools include support for remote Model Context Protocol servers, image generation, a Code Interpreter and improvements to file search. They are available across GPT-4o series, GPT-4.1 series and OpenAI o-series reasoning models.
OpenAI on Wednesday also said it will buy io, an AI device startup founded by iPhone and Apple Watch designer Jony Ive, for $6.5 billion.
Esther Shittu is an Informa TechTarget news writer and podcast host covering artificial intelligence software and systems.
