Mistral Medium 3.5 inside Vibe CLI — when the open-weight default becomes the lab's own production choice
Mistral made Medium 3.5 the default in Le Chat and replaced Devstral 2 in Vibe CLI this week. Open-weight model coverage usually stops at benchmark scores; this commitment moves the open-weight tier from "option for cost-sensitive workloads" to "production default at the lab itself."
Mistral Medium 3.5 — 77.6% SWE-Bench Verified — is now the default in Le Chat conversations and Vibe CLI's underlying coding model. Open-weight model releases historically get benchmark coverage; making the same model the default in the lab's own consumer assistant and canonical coding agent says something stronger about how Mistral views production-readiness of the open-weight tier.
The Apache-2.0 stack just got tighter
Mistral Large 3 and Mistral Small 4 already ship under Apache 2.0 — the most permissive open-weight license in the frontier-adjacent tier. Adding Voxtral TTS as the multilingual speech layer means Mistral's product surface now spans LLM chat, coding agent, and speech generation — all open-weight, all routed through one operational layer.
Why Vibe CLI is the procurement-relevant signal
Vibe CLI is the canonical Mistral coding agent — the workflow shape that competes with Claude Code and OpenAI Codex. Until this week, Vibe CLI ran Devstral 2 (a coding-specialist fine-tune). Replacing that with Medium 3.5 — the unified Magistral-reasoning + Devstral-coding model — says Mistral believes the unified base model with strong agentic-tool capabilities beats the specialist fine-tune in production. That's a meaningful internal signal for what the open-weight tier can do.
What this does for European procurement
For European sovereign-AI buyers, the open-weight + EU-headquartered stack is now: Mistral Medium 3.5 (text + code + reasoning), Voxtral (TTS), plus DeepSeek/Qwen/Llama variants for diversity. Combined with the Code of Practice on AI content labelling and the Digital Omnibus amendments, the procurement story for August 2 compliance just has more credible Tier-1 European options than at any point in the AI Act's enforcement runway.
The competitive frame against proprietary mid-tier
Gemini 3.5 Flash at $1.50/$9 per million tokens. GPT-5.5 at $1.50/$9. Claude Sonnet 4.5 at $3/$15. Mistral Medium 3.5 at open weights with permissive licensing. The procurement question used to be "which proprietary mid-tier model wins"; it's increasingly "do I need a proprietary mid-tier model at all when the open-weight option is at parity and I can host it myself." The Vibe CLI default is Mistral's bet that the answer is increasingly no.
Mistral AI — Mistral AI Newsroom → · Releasebot — Mistral Release Notes - June 2026 Latest Updates →