Mistral AI Devstral 2 and Mistral Vibe CLI Release

Mistral AI has announced the release of Devstral 2 and Devstral Small 2, a new family of open-weight coding models designed for high-performance code agents. These models are paired with Mistral Vibe, a native open-source CLI tool for end-to-end code automation.

Devstral 2: High-Efficiency Coding Models

Devstral 2 establishes a new open-weight state-of-the-art for code agents, achieving a 72.2% score on SWE-bench Verified. The model family consists of two sizes: Devstral 2 (123B parameters) and Devstral Small 2 (24B parameters), both supporting a 256K context window.

Model Specifications and Licensing

  • Devstral 2 (123B): A dense transformer released under a modified MIT license. It is optimized for data center GPUs, requiring a minimum of 4 H100-class GPUs for deployment.
  • Devstral Small 2 (24B): Released under the Apache 2.0 license. This model is designed for local deployment on consumer hardware, including single-GPU NVIDIA systems (DGX Spark, GeForce RTX) and CPU-only configurations.

Performance and Efficiency

Devstral 2 is significantly more compact than competitors like DeepSeek V3.2 (5x smaller) and Kimi K2 (8x smaller). Despite this reduced size, it maintains high performance, while Devstral Small 2 is 28x smaller than DeepSeek V3.2 and 41x smaller than Kimi K2.

In human evaluations conducted by an independent provider using Cline, Devstral 2 demonstrated a 42.8% win rate against DeepSeek V3.2 (with a 28.6% loss rate). While Claude Sonnet 4.5 remains preferred over Devstral 2, the model is described as being up to 7x more cost-efficient than Claude Sonnet for real-world tasks.

Production-Grade Workflow Capabilities

Devstral 2 is built for complex software engineering tasks, including:

  • Multi-file Orchestration: The ability to explore codebases and orchestrate changes across multiple files while maintaining architecture-level context.
  • Dependency Tracking: Tracking framework dependencies and detecting failures to retry with corrections.
  • Enterprise Customization: Support for custom fine-tuning to prioritize specific languages or optimize for large enterprise codebases.
  • Multimodal Support: Devstral Small 2 supports image inputs, enabling the power of multimodal agents.

Mistral Vibe CLI: Autonomous Code Automation

Mistral Vibe CLI is an open-source command-line assistant (Apache 2.0 license) powered by Devstral. It allows developers to integrate AI-driven code modification and execution directly into their terminal or IDE via the Agent Communication Protocol.

Key Features of Vibe CLI

  • Project-Aware Context: The CLI automatically scans file structures and Git status to provide relevant context for the AI.
  • Smart References: Supports @ autocomplete for files, shell command execution via !, and slash commands for configuration.
  • Architecture-Level Reasoning: By understanding the entire codebase, Vibe CLI can perform multi-file orchestration to reduce PR cycle times.
  • Customization: Includes persistent history, customizable themes, and a config.toml for configuring local models and providers.
  • Control: Users can toggle auto-approval for tool execution and manage tool permissions.

Deployment and Pricing

Devstral 2 is currently available for free via the Mistral AI API. Following the free period, pricing will be set as follows:

Model Input Price (per 1M tokens) Output Price (per 1M tokens)
Devstral 2 $0.40 $2.00
Devstral Small 2 $0.10 $0.30

Devstral 2 is available on build.nvidia.com and will soon receive NVIDIA NIM support. For optimal performance, Mistral recommends a temperature of 0.2.

Sources

Related

  • Dispatch
    Mistral AI Devstral Models Release

    Mistral AI has released Devstral Medium and Devstral Small 1.1, providing state-of-the-art agentic coding capabilities with a focus on generalization across different prompts and agentic scaffolds.

  • Dispatch
    Mistral AI Devstral Release

    Mistral AI has released Devstral, an agentic LLM for software engineering that outperforms open-source models on the SWE-Bench Verified benchmark and is available under the Apache 2.0 license.

  • Dispatch
    Mistral Vibe 2.0 Release Notes

    Mistral AI has released Mistral Vibe 2.0, a terminal-native coding agent powered by the Devstral 2 model family that introduces custom subagents, slash-command skills, and unified agent modes.

  • Dispatch
    Mistral AI Codestral 25.08 and Enterprise Coding Stack Release

    Mistral AI has released Codestral 25.08 and a comprehensive enterprise coding stack featuring integrated completion, semantic search, and agentic workflows with flexible deployment options.

  • Dispatch
    Mistral AI September 2024 Release

    Mistral AI has introduced a free tier for la Plateforme, significant price reductions across its model family, the Mistral Small v24.09 model, and free vision capabilities via Pixtral 12B on le Chat.