Mistral AI releases Vibe remote coding agent and 128 billion parameter model Medium 3.5

AIMPACT News, May 3rd (UTC+8), Mistral AI released the Vibe remote coding agent and the public preview version of Mistral Medium 3.5. Vibe is a CLI-accessible coding agent capable of completing software tasks such as code writing, refactoring, test generation, and CI failure investigation, supporting cloud backend operation, allowing users to handle long-duration tasks even when away from their computers, and supporting multiple agents in parallel. Each agent session runs in an isolated sandbox, and upon completion, can create a Pull Request on GitHub and notify the user. Mistral Medium 3.5 is a dense model with 128 billion parameters, marking its first flagship integrated model, with a 256k-token context window, supporting instruction following, reasoning, and coding tasks. The model is trained from scratch with a visual encoder to handle variable image sizes. On the SWE-Bench validation set, it scored 77.6%, surpassing Devstral 2 and Qwen3.5 397B A17B, and has become the default model for Vibe and Le Chat. Inference effort can now be configured upon request.