Mistral launches Medium 3.5 flagship and 'Vibe' agent with VS Code integration

Mistral AI released Mistral Medium 3.5, a 128-billion-parameter dense flagship that unifies instruction-following, reasoning, and coding in a single model, with a deployment story tuned to enterprise control: it can be self-hosted on as few as four GPUs. The dense (non-MoE) design and modest hardware footprint are deliberate differentiators for customers wary of cloud lock-in.

Mistral paired the model with 'Vibe,' an AI agent for work and code featuring a Work Mode for long-running, multi-step tasks and a Code Mode for remote coding, plus a new VS Code extension to meet developers in their existing workflow. The agent framing slots Mistral into the same long-horizon-agent race as Claude Code, Codex, and Grok Build.

Competitively, Mistral's pitch remains European sovereignty plus self-hostability — a hedge against US hyperscaler dependence that resonates with regulated industries and EU buyers. The four-GPU self-host claim is the headline practical detail; it makes Medium 3.5 viable for on-prem deployments that can't or won't send data to a hosted API.

Caveats: 128B dense is large to serve at scale, and 'four GPUs' likely assumes high-end accelerators and aggressive quantization — real throughput will vary. What to watch: independent benchmarks versus comparable open and hosted models, VS Code extension adoption, and whether Vibe's Work/Code modes meaningfully differentiate from incumbent coding agents.