AI startup Mistral revealed its new AI model for coding, Devstral, in an announcement this week. Developed in collaboration with AI firm All Hands AI, Devstral is designed specifically for programming tasks and is publicly available under an Apache 2.0 license, making its commercial use unrestricted.
Mistral asserts that Devstral has already outperformed other leading open-source coding models like Google’s Gemma 3 27B and DeepSeek’s V3 according to the SWE-Bench Verified test, an evaluation tool focused on assessing advanced code-writing capabilities.
In a blog post, Mistral highlighted Devstral’s strengths, noting its capability to interact efficiently with coding repositories, handle multi-file project edits smoothly, and power software engineering AI agents. The model works seamlessly with popular software agent platforms like OpenHands or SWE-Agent, which provide standardized integration between Devstral and testing frameworks. Notably, Devstral is efficient enough to operate on modest hardware setups, including a single Nvidia RTX 4090 graphics card or a Mac equipped with 32GB of RAM, paving the way for flexible local usage.
The announcement comes amid a surge of interest in AI-driven coding assistants, with many prominent technology companies launching dedicated coding models of their own. Recently, JetBrains, known for its widely-used developer tools, introduced its first public AI model aimed at coding tasks. Google, Windsurf, and OpenAI have also recently unveiled new models specifically optimized for programming applications, reflecting the growing demand in this space.
Despite significant industry enthusiasm, AI-generated code still has notable limitations, frequently introducing errors and security weaknesses due to challenges related to understanding complex programming logic. Nonetheless, the compelling promise of increased productivity has convinced many developers and organizations to adopt these specialized tools rapidly. A recent survey revealed that approximately 76% of software developers reported either currently using or planning to implement AI coding tools in their workflows.
This is not Mistral’s first foray into programming-focused AI. Previously, the company released Codestral, a generative AI model for coding tasks. Unlike Devstral, however, Codestral was not licensed for commercial internal uses by corporate teams, restricting its adoption in business-oriented environments.
Interested users can now access Devstral either by downloading it from platforms like Hugging Face or by interacting with it via Mistral’s API service, which charges $0.1 per million input tokens and $0.3 per million output tokens—a token being a basic data unit that AI models process. Devstral, with a parameter count of around 24 billion, is positioned competitively but still moderately sized compared to larger state-of-the-art models. Mistral added that a more robust, powerful coding model is already in development and will be available shortly.
Founded in 2023, Mistral has rapidly established itself as a major player in the frontier AI market. Backed by venture capital firms such as General Catalyst, the company has secured over €1.1 billion (approximately $1.24 billion) in total funding. It serves prominent clients including BNP Paribas, AXA, and Mirakl.
Devstral marks the company’s third notable product launch this month. Earlier in May, Mistral introduced the Mistral Medium 3 general-purpose AI model, as well as Le Chat Enterprise—a corporate chatbot solution capable of integrating with select enterprise tools, such as Gmail, Google Drive, and SharePoint.
