Mistral bets big on “vibe coding” with new autonomous software-engineering agent

On Tuesday, the French AI startup Mistral AI made headlines with the launch of Devstral 2, an impressive coding model boasting 123 billion parameters. This groundbreaking open-weights model is designed to function as part of an autonomous software engineering agent, achieving a remarkable score of 72.2 percent on the SWE-bench Verified benchmark. This score places it among the elite open-weights models, demonstrating its capability to address genuine issues found on GitHub. In addition to the AI model, Mistral introduced an innovative development application known as Mistral Vibe. This command line interface (CLI) is comparable to tools like Claude Code, OpenAI Codex, and Gemini CLI, allowing developers to engage with the Devstral models directly from their terminals. Mistral Vibe can scan file structures and analyze Git status, effectively maintaining context throughout an entire project. It can autonomously execute shell commands and make changes across multiple files, streamlining the development process. While AI benchmarks should be approached with caution, industry insiders have indicated that performance on SWE-bench Verified is closely monitored by major AI companies. This benchmark presents AI models with 500 real software engineering challenges sourced from popular Python GitHub repositories. The AI must comprehend the issue description, navigate the codebase, and produce a functional patch that successfully passes unit tests. Though some researchers have pointed out that about 90 percent of the tasks in the benchmark consist of relatively straightforward bug fixes that seasoned engineers could resolve in under an hour, it remains one of the few standardized methods for evaluating coding models. Alongside its flagship model, Mistral also unveiled Devstral Small 2, a 24 billion parameter variant that achieved a score of 68 percent on the same benchmark. This smaller model is capable of running locally on consumer hardware, such as laptops, without requiring an internet connection. Both models support a substantial context window of 256,000 tokens, enabling them to process moderately large codebases, although perceptions of size may vary based on project complexity. Mistral has released Devstral 2 under a modified MIT license and Devstral Small 2 under the more permissive Apache 2.0 license, furthering its commitment to the open-source community.