Mistral AI Launches New Coding Models
ā¼ Summary
– Mistral AI launched Devstral 2, a new AI model for coding, and Mistral Vibe, a CLI tool, to compete with larger rivals in the AI coding space.
– The new tools emphasize context awareness for business workflows, with Vibe CLI using persistent history and scanning file structures to inform its behavior.
– Devstral 2 is a large, demanding model requiring significant GPU resources, while a smaller version, Devstral Small, is designed for local deployment on consumer hardware.
– The models have different open-source licenses (modified MIT and Apache 2.0) and pricing structures, with Devstral 2 currently free via API before transitioning to a paid model.
– Mistral AI, valued at approximately $13.8 billion, has partnered with agent tools Kilo Code and Cline for Devstral 2’s release and made Vibe CLI available as an IDE extension.
Mistral AI has introduced a new generation of its specialized coding model, Devstral 2, alongside a novel command-line tool named Mistral Vibe. This strategic launch underscores the French startup’s ambition to compete more directly with larger, well-funded AI laboratories and other prominent coding-focused large language models. The announcement builds on the recent release of the Mistral 3 model family, signaling a clear push to narrow the gap with its rivals.
A significant part of this initiative involves entering the competitive arena of “vibe-coding,” a trend popularized by tools like Cursor and Supabase. The Mistral Vibe CLI is designed to streamline code automation using natural language commands. It provides capabilities for file manipulation, searching through codebases, managing version control, and executing system commands directly from the terminal.
The company emphasizes the critical importance of context awareness for practical business applications. Similar to its conversational AI, Le Chat, which retains memory of past interactions, the Vibe CLI features a persistent history. It can also intelligently scan file structures and check Git repository statuses to build a rich understanding of the developer’s environment, thereby informing more accurate and relevant automated actions.
This focus on supporting robust, production-level workflows is reflected in the technical specifications of Devstral 2. The model is substantial, with 123 billion parameters, and requires significant computational power for deployment, such as a minimum of four H100 GPUs or equivalent hardware. To accommodate different needs, Mistral also offers Devstral Small, a more compact version with 24 billion parameters that can be run locally on consumer-grade hardware.
The two models differ in their open-source licensing approaches. Devstral 2 is released under a modified MIT license, while Devstral Small utilizes the Apache 2.0 license. Their pricing structures also vary. Currently, Devstral 2 is accessible for free via Mistral’s API. After this introductory period, API costs will be set at $0.40 per million input tokens and $2.00 per million output tokens. The smaller model, Devstral Small, will be priced at $0.10 and $0.30 per million tokens for input and output, respectively.
To facilitate adoption, Mistral has collaborated with agent tools Kilo Code and Cline for the distribution of Devstral 2. The Mistral Vibe CLI is available as an extension within the Zed integrated development environment (IDE), allowing developers to use it directly in their coding workspace.
As Europe’s leading AI lab, Mistral AI holds a valuation of approximately ā¬11.7 billion. This valuation followed a substantial Series C funding round where Dutch semiconductor giant ASML invested ā¬1.3 billion, demonstrating strong investor confidence in the company’s trajectory and technological vision.
(Source: TechCrunch)
