Mistral Medium 3.5: coding agents move to the cloud
On May 2, 2026, Mistral AI moved its Medium 3.5 model into public preview and shipped remote agents for its Vibe coding platform together with an agentic Work Mode in Le Chat. The shift for coding workflows is from “local with the occasional cloud run” to “cloud as default, local when needed” — backed by a self-reported 77.6% on SWE-Bench Verified, which puts the model at the front of dense open-weight contenders.
What concretely changed
- Mistral Medium 3.5 is now the default in Le Chat and Vibe — 128B dense parameters, 256K context window, open weights under a modified MIT license.
- 77.6% SWE-Bench Verified — Mistral places this ahead of Devstral 2 and several larger MoE models. The τ³-Telecom score is 91.4.
- Remote Agents in Vibe run coding sessions in isolated cloud sandboxes. A local session can be “teleported” — history and task state move to the cloud, and multiple agents can run in parallel.
- Work Mode in Le Chat wires the agent into email, calendar, documents, Jira and Slack — with a visible tool trace and explicit approval before sensitive actions.
What used to be
Until now, Mistral Vibe was a locally running coding companion — comparable to Claude Code or the Codex CLI: you start the CLI, the agent works on your machine, your terminal keeps the session alive. Long-running tasks (refactors, migrations, sweeping test generation) tied up the local machine — and the session died the moment the laptop closed.
Le Chat stayed mostly reactive as a chat interface. The “Deep Research” mode from summer 2025 enabled longer research runs but no autonomous orchestration of multiple tools with write access into connected systems.
What now applies
1. Coding sessions can migrate to the cloud. Anyone working locally in Vibe can lift a running session into a cloud sandbox with a single command. Mistral calls it “teleporting” — history and state are preserved, the agent keeps running asynchronously, and multiple agents can work on different tasks in parallel. Tool connectors for GitHub, Linear, Jira, Sentry, Slack and Teams are built in; PRs can be opened straight from the agent.
2. Le Chat has an agentic Work Mode. It plans across several tools in one go — for example: “read the last three Jira tickets, draft a Slack status update from them, put a meeting on tomorrow’s calendar for 10 a.m.” All tool calls and reasoning steps are visible, sensitive actions need explicit approval. That moves Le Chat closer to the “enterprise assistant” position that ChatGPT Team and Claude for Work have occupied since early 2026.
3. Medium 3.5 becomes the default model. Le Chat and Vibe now route to Medium 3.5; the model is also available with open weights under a modified MIT license on Hugging Face. The 77.6% SWE-Bench Verified score is self-reported — comparable numbers usually only show up on much larger reasoning models. In Le Chat, reasoning effort can be tuned per API request, which saves tokens inside agent loops.
Context
The real story behind this release isn’t the model — it’s the infrastructure shift. Mistral is following a pattern Anthropic established with Skills/Hooks and OpenAI with Operator — but as a European vendor with open model weights. For teams that, for compliance reasons, can’t route code through US hyperscalers, that’s a credible alternative.
It’s also notable that Mistral pulls coding and office workflows into the same tool layer. Vibe speaks GitHub and Linear, Le Chat speaks email and calendar — both run on the same model. If you’re already looking for one central AI platform for several teams, that’s one fewer set of point integrations to maintain.
What you can do now
If you’re evaluating coding agents: Vibe with remote agents fits long, asynchronous tasks — comparable to “Claude Code in the background.” Open weights additionally allow self-hosting; that’s not on the table with the US competitors.
If your team is already on Le Chat: Work Mode is worth testing against your existing workflow automation. The visible tool trace with an approval step is the trustworthy variant — it gets fragile fast if approvals get clicked through without a glance.
If you care about the comparison to Claude or GPT: Plan your own eval on your data. SWE-Bench numbers are self-reported, and the model-to-model jumps in synthetic benchmarks often look larger than they feel in daily work.
Entdecke mehr
Headless without an API bill — how do you reach the best AI models for automation in 2026?
Provider comparison mid-2026: who has a headless mode, whose subscription still covers it — and why BYOK is the most stable foundation.
GlossarAider
Aider is an open-source AI pair programmer for the command line that edits code directly in the Git repository and commits every change automatically.
LexikonComparing AI models — who builds what and how to choose
The major model families in 2026 at a glance. Who builds Claude, GPT, Gemini, Llama, Mistral, DeepSeek, Qwen — and which model to pick when.