What Devstral 2 Brings
French company Mistral AI has launched Devstral 2, a new family of models designed primarily for coding. This model has 123 billion parameters and operates as a dense transformer with a context window of up to 256,000 tokens. It scores 72.2% on the SWE-bench Verified benchmark, ranking it among the best open models. It is available under a modified MIT license and is intended to accelerate the distribution of intelligence.
Devstral 2 can navigate entire codebases, manage changes across multiple files simultaneously, and maintain context at the architectural level. It tracks framework dependencies, identifies errors, and iterates on fixes. It handles tasks such as bug fixes and the modernization of legacy systems. It can be customized for specific languages or large enterprise codebases.
Comparison with other models: Devstral 2 has an advantage over DeepSeek V3.2, with a 42.8% win rate versus a 28.6% loss rate in human evaluations conducted by an independent annotation provider, with tasks managed by Cline. However, Claude Sonnet 4.5 remains the preferred model. Devstral 2 is 5x smaller than DeepSeek V3.2 and 8x smaller than Kimi K2, yet delivers comparable performance.
Smaller Variant: Devstral Small 2
In addition to the main model, Devstral Small 2 comes with 24 billion parameters. It scores 68.0% on SWE-bench Verified and ranks alongside models up to five times larger. It runs on standard hardware, such as consumer GPUs, or even on a CPU alone without dedicated graphics. It has the same 256,000-token context window and supports image inputs for multimodal agents. It is licensed under Apache 2.0.
Devstral Small 2 is 28x smaller than DeepSeek V3.2 and 41x smaller than Kimi K2. Its size enables fast inference, short feedback loops, and easy customization.
Mistral Vibe CLI: An Assistant Directly in the Terminal
At the same time, Mistral AI introduced Mistral Vibe CLI, an open-source tool for automating code-related tasks through natural language. It works in the terminal or integrates into IDEs via the Agent Communication Protocol. It is licensed under Apache 2.0 and powered by Devstral.
This CLI explores, edits, and applies changes to the codebase. It offers an interactive chat with tools for file manipulation, code search, version control, and command execution. Key features include automatic scanning of the file structure and Git status for relevant context, smart references with autocomplete using @ for files, ! for shell commands, and slash commands for configuration.
It supports orchestration across multiple files, understands the entire codebase, and enables architectural-level reasoning, potentially cutting the pull request cycle in half. It has persistent history, autocomplete, and customizable themes. It can be run programmatically for scripts, toggle automatic tool approval, configure local models through config.toml, and manage permissions.
Mistral Vibe CLI is available as an extension in Zed, allowing it to be used directly in the IDE.
Availability and Pricing
Devstral 2 is currently available for free via the API at console.mistral.ai. After the free period, pricing will be $0.40 per million input tokens (CZK 9) and $2.00 per million output tokens (CZK 47). For Devstral Small 2, the prices will be $0.10 for input (CZK 2) and $0.30 for output (CZK 7).
Mistral is collaborating with Kilo Code and Cline to deploy Devstral 2. Deploying Devstral 2 requires at least four H100 GPUs, and it can be tested at build.nvidia.com. Devstral Small 2 runs on a single GPU, such as a DGX Spark or GeForce RTX, with NVIDIA NIM support coming soon.
Mistral AI has a valuation of €11.7 billion (approximately CZK 300 billion) following a €1.3 billion investment from holding company ASML in September.
Devstral 2 had a successful launch, exceeding 17 billion tokens within the first 24 hours. According to Cline, its tool-calling success rate is on par with the best closed models. Kilo Code describes it as a model that operates at scale with low costs.
