On April 29, Mistral AI released Mistral Medium 3.5, with the 128B dense model scoring 77.6% on SWE-Bench Verified, runnable on four GPUs, with 256K context. The Vibe remote coding agent was also launched simultaneously, supporting cloud multi-agent parallelism, cross-environment state preservation, and automatic GitHub PR submission. The model uses a custom vision encoder, and Vibe is built on Mistral Studio Workflow orchestration layer, providing developers with a referenceable Agent architecture paradigm.