2026-04-29
Mistral releases Medium 3.5 with 256K context, 77.6% SWE-Bench Verified
Mistral Medium 3.5: 128B dense, 256K context, 77.6% SWE-Bench Verified. Vibe gets cloud remote agents; Le Chat gets a Work Mode.
Mistral released Medium 3.5 — a 128B-parameter dense model with a 256K context window — and reports 77.6% on SWE-Bench Verified. Alongside the model, Vibe (Mistral’s CLI) gains cloud-hosted remote agents that run in isolated GPU sandboxes with parallel sessions and state preservation across local↔cloud transitions. Le Chat also adds a “Work Mode” for multi-step, cross-tool workflows with human-in-the-loop approval.
Practitioner note
A 77.6% SWE-Bench Verified score from an open-weight-friendly vendor (Mistral has consistently released open-weights at the lower tiers) is a credible alternative to Claude/GPT for self-hosted agentic-coding setups.
Why this matters for you specifically: if you’re already running a LiteLLM-routed local stack (which this project is set up for), Mistral Medium 3.5 — once weights or hosted endpoints settle — gives you a serious option to point your local agents at, beyond the Qwen series. The 256K context is more than enough for most agentic-coding sessions.