13 December 2025

French unicorn Mistral AI has launched Devstral 2, a new coding‑focused language model, and Vibe, a command‑line interface that turns natural language into code. Devstral 2 weighs 123 billion parameters and requires four H100 GPUs to deploy, while the smaller Devstral Small (24 billion parameters) can run locally . The models are free to use initially and will later cost $0.40/$2.00 per million input/output tokens for Devstral 2 and $0.10/$0.30 for Devstral Small. Vibe CLI features context awareness and persistent history, scanning file structures and Git statuses to inform its actions. Mistral’s release underscores Europe’s growing AI ambitions and offers developers a new set of tools for automating coding tasks.

Article

While much of the AI world watches U.S. and Chinese giants, European players like Mistral AI are rapidly building their own ecosystems. On 9 December 2025, Mistral announced Devstral 2, a generational upgrade to its code‑generation models . The company also unveiled Devstral Small, a lighter version, and Mistral Vibe, a command‑line interface inspired by the “vibe‑coding” movement .

Devstral 2 is designed for large‑scale deployments. At 123 billion parameters, it falls between OpenAI’s GPT‑4 and Anthropic’s Claude 3 in size. Deploying it requires at least four NVIDIA H100 GPUs or equivalent hardware . The model is available under a modified MIT license and is free to use during an introductory period; after that, the API will charge $0.40 per million input tokens and $2.00 per million output tokens. Devstral Small, by comparison, weighs 24 billion parameters, uses an Apache 2.0 license and will cost $0.10/$0.30 per million tokens . The smaller model is optimized to run locally on consumer hardware, giving developers more control over their code and data.

Mistral didn’t stop at models. It introduced Vibe, a command‑line interface (CLI) that leverages natural language to automate coding tasks. Vibe features persistent history and can scan file structures and Git statuses to build context. This context awareness allows Vibe to execute multi‑step commands—search files, modify code, commit to version control—without leaving the terminal. The idea is to empower developers to “code by vibe,” turning high‑level intentions into working code.

The vibe‑coding trend has gained traction thanks to startups like Cursor and Supabase. It emphasizes a conversational workflow where developers state their goals and let the AI handle the details. Vibe aims to mainstream this concept by integrating with the Zed code editor and offering an extension for easy adoptiont. Mistral also partnered with agent‑tools providers Kilo Code and Cline to deliver Devstral 2 via API.

Mistral’s announcement comes on the heels of its Mistral 3 family of open‑weight models released earlier this year. Together, these products signal the company’s ambition to challenge larger AI labs on multiple fronts—open‑source models, enterprise APIs and developer tooling. Mistral is already valued at €11.7 billion following a Series C round led by semiconductor giant ASML, which invested €1.3 billion. That backing gives the French startup resources to compete globally while maintaining a European identity.

For entrepreneurs, the arrival of Devstral and Vibe offers new possibilities. Startups can embed Devstral’s coding capabilities into their products, automating boilerplate generation, code reviews and documentation. The open licensing makes it easier to integrate the models without onerous restrictions. Meanwhile, Vibe CLI’s context awareness could streamline internal workflows by turning natural language instructions into code changes, saving precious developer hours.

There are caveats. Large models like Devstral 2 consume significant compute and energy. Only companies with access to powerful GPUs will be able to deploy the full model. The smaller Devstral, while more accessible, still requires careful tuning and may not match the performance of larger models on complex tasks. Pricing beyond the free period could also become a significant expense for startups.

Mistral AI’s dual launch of Devstral 2 and Vibe CLI illustrates that AI innovation isn’t confined to Silicon Valley or Shenzhen. By providing powerful coding models under permissive licenses and pairing them with context‑aware tooling, the French startup invites developers worldwide to embrace a new, conversational approach to programming. Early adopters who integrate these tools can accelerate product development and contribute to a more diverse AI landscape.

Subscribe to our newsletter to stay up to date with the latest breakthroughs in AI, Business, Technology, and Mindset.