Mistral drops Devstral 2 coding model and Vibe CLI

Devstral 2 is Mistral AI's new open-weight coding family, built specifically for multi-file repo edits, bug fixing, and autonomous software engineering. Alongside the models, Mistral also dropped Mistral Vibe CLI, a native terminal-based agent that automates your development workflow.

The family comes in two sizes: the flagship Devstral 2 with 123B parameters, and Devstral Small 2 with 24B parameters. Both feature a massive 256K context window, letting them ingest entire enterprise codebases at once.

What it does and how it performs

Devstral 2 is designed for production codebases rather than simple single-file completions. It tracks dependencies across multiple files, reads Git status, catches runtime errors, and autonomously loops to fix broken builds.

On the SWE-bench Verified benchmark, the 123B model hits 72.2%, while Mistral claims up to 7x the cost efficiency of Claude Sonnet on real-world engineering tasks. The 24B Devstral Small 2 scores 68.0%, effectively punching way above its weight class.

In human evals run through the Cline agent scaffold, Devstral 2 beat DeepSeek V3.2 with a 42.8% win rate against 28.6%. Cline noted that its tool-calling reliability matches top-tier proprietary models. However, human evaluators still clearly favored Claude Sonnet 4.5, so the frontier closed-source moat isn't entirely gone.

The Mistral Vibe CLI rounds out the package. Running under Apache 2.0, it gives your terminal native agentic powers: @ to reference files with autocomplete, ! to execute shell commands, and persistent memory across sessions. It works in the terminal or inside Zed via an official extension.

Specs, licenses, and hardware

The two models target different setups.

Devstral 2 (123B) is released under a modified MIT license. Running it on-prem requires serious iron: at least four H100-class GPUs. You can also test it via build.nvidia.com or Mistral's API, which is currently free during its launch phase.

Devstral Small 2 (24B) is released under Apache 2.0. It supports multimodal image inputs and runs locally on consumer hardware, including single GeForce RTX GPUs and even CPU-only machines for private, offline work.

Once the free API window closes, Devstral 2 will cost $0.40 per million input tokens and $2.00 per million output tokens. Devstral Small 2 runs at a dirt-cheap $0.10/$0.30 per million tokens.

Why it matters

We think Devstral 2 strikes the right balance between model size and agentic capability. At 123B, it is far smaller than DeepSeek V3.2 or Kimi K2, yet it delivers the tool-calling precision necessary to actually edit real code.

If you want a private local code assistant, Devstral Small 2 running with Vibe CLI is worth spinning up immediately. For teams looking to cut down massive Anthropic bills on daily PR workflows, the 123B model offers a compelling, self-hostable alternative.