<RETURN_TO_BASE

Mistral AI Unveils Devstral 2 and Vibe CLI

Mistral AI introduces Devstral 2 coding models and Vibe CLI for enhanced terminal development.

Cutting-edge Coding Models by Mistral AI

Mistral AI has introduced Devstral 2, a next-generation coding model family tailored for software engineering agents, along with Mistral Vibe CLI, an open-source command line coding assistant that operates within the terminal or IDEs supporting the Agent Communication Protocol.

Specifications and Performance Metrics

Devstral 2 is a 123B parameter dense transformer with a 256K token context window. It achieves a score of 72.2% on SWE-bench Verified, placing it among the top open-weight models for software engineering tasks, and is available under a modified MIT license. In contrast, Devstral Small 2 has 24B parameters and also utilizes a 256K context window, scoring 68.0% on SWE-bench Verified, and is published under the Apache 2.0 license, fostering adoption for production use.

Both models are deemed open source and permissively licensed, showcasing their efficiency: Devstral 2 is reported to be up to 7 times more cost-effective than Claude Sonnet for similar coding quality tasks, vital for continuous agent workflows. Notably, they are significantly smaller than leading systems, being 5 and 28 times smaller than DeepSeek V3.2.

Ideal for Production-grade Coding Workflows

Designed to cater to complex software engineering tasks, Devstral 2 can adeptly explore code repositories, track dependencies, and orchestrate edits across multiple files while retaining architectural context. It additionally supports tasks such as bug fixing and modernization of legacy systems.

The model can be fine-tuned for specific programming languages or extensive enterprise codebases. Meanwhile, Devstral Small 2 serves a similar purpose but is optimized for local deployments and private runtimes.

Evaluations against Competitors

Human evaluations revealed that Devstral 2 outperformed DeepSeek V3.2 and Claude Sonnet 4.5 with a win rate of 42.8%. This competitive edge underscores its potential in real-world coding scenarios.

Introducing Mistral Vibe CLI

The Mistral Vibe CLI serves as a terminal-native coding agent, developed in Python and powered by the Devstral models. It provides functionalities such as exploring and modifying codebases using natural language directly in the terminal or IDEs like Zed, where it is available as an extension.

Key Features of Vibe CLI

  • Project-Aware Context: Scans file structures for a comprehensive view of the repository.
  • Smart References: Includes features like @ autocompletion for files and shell commands.
  • Multi-File Orchestration: Coordinates architecture-level changes across the entire codebase.
  • Persistent History: Offers autocompletion and customizable themes for daily terminal usage.

Developers can configure Vibe CLI using a simple config.toml file to connect to Devstral 2 via the Mistral API, integrating programmatic capabilities and granular permissions.

Conclusion: Key Takeaways

  1. Devstral 2 is a 123B parameter model with 72.2% SWE-bench Verified, released under a modified MIT license.
  2. Devstral Small 2 boasts 24B parameters with applications under the Apache 2.0 license.
  3. Both models optimize agentic coding workflows, enabling robust codebase exploration and multi-file edits.
  4. Mistral Vibe CLI enhances development with a chat-style interface, focusing on project context and multi-file coordination.

More technical details can be found on the Mistral AI website.

🇷🇺

Сменить язык

Читать эту статью на русском

Переключить на Русский