Mistral Releases Open-Source Coder 'Devstral 2'
Mistral has launched Devstral 2, a 123B parameter agentic coding model that scored 72.2% on the SWE-bench. It's paired with an open-source Vibe CLI and priced aggressively at a reported 7x cheaper than Claude Sonnet, intensifying the price war in the agentic coding market. A smaller 24B version is designed to run locally on consumer hardware.
Devstral 2 is a 123B-parameter dense transformer, not a Mixture-of-Experts (MoE) model, and supports a 256K context window. This dense architecture is a notable choice, aiming for high performance-to-efficiency without the routing complexity of MoE systems. The model is also multimodal, capable of processing images alongside text and code. The 72.2% score was achieved on SWE-bench Verified, a benchmark that uses real-world GitHub issues to evaluate a model's ability to generate patches for software problems. This places it among the top open-weight models, though behind the latest proprietary models like Claude Opus 4.6, which has scored 79.20% on the same benchmark. The accompanying Vibe CLI is an open-source (Apache 2.0) agent that runs directly in the terminal, allowing it to read your file structure and Git status for project-aware context. It can execute file manipulations, run shell commands, and manage version control, integrating with IDEs like Zed and Kilo Code via the Agent Communication Protocol. The smaller 24B parameter "Devstral Small 2" model is licensed under Apache 2.0 and achieves a 68.0% on SWE-bench Verified. It is designed to run on a single consumer GPU, requiring 24GB of VRAM (like an RTX 3090/4090) or a Mac with at least 32GB of RAM for local inference. Running the full 123B model requires a significant hardware investment, needing a minimum of 72GB of VRAM just to load with a small context. For production use with its full 256K context window, a multi-GPU setup such as four H100s is recommended. The pricing for the 123B model is set at $0.40 per million input tokens and $2.00 per million output tokens. This is roughly 7x cheaper than the standard rate for Anthropic's Claude Sonnet 4.6, which is priced at $3 per million input and $15 per million output tokens. This release comes as Mistral's valuation has surged, reaching approximately €12 billion ($14 billion) after a €2 billion investment round in September 2025. The funding positions Mistral as a key European competitor to U.S.-based AI giants.
