Devstral 2 123B

Devstral the listed price is a 123-billion-parameter agentic language model purpose-built for software engineering. Unlike general-purpose assistants, it is engineered to explore codebases, edit multiple files, and drive autonomous coding agents—making it particularly effective for tasks requiring sustained tool use and multi-step reasoning across a repository. The model ships in FP8 precision, balancing memory efficiency with the fidelity needed for complex code generation and comprehension. Its design centers on agentic coding, where the model can plan, execute, and iterate on code changes with minimal human intervention, which positions it as a strong foundation for building AI-powered development pipelines.

The model is released as an Instruct variant, meaning it undergoes instruction fine-tuning to sharpen responsiveness to human-directed tasks. Benchmark results underscore its capabilities: Devstral the listed price scores 72.2% on SWE-bench Verified and 61.3% on SWE-bench Multilingual, alongside a Terminal Bench the listed price score of 32.6%. These numbers place it competitively against larger models in real-world software engineering evaluations. The model carries a Modified MIT license and is available through multiple deployment paths including NVIDIA NIM, Ollama, and Hugging Face, making it accessible for both research and production use. For enterprises needing expanded context, domain-specific customization, or specialized capabilities, the creators invite direct outreach, suggesting a pathway toward tailored deployments beyond the base model.