Mistral Launches Vibe Remote Agents — Async Cloud Coding With Open-Weight Models
Mistral's Vibe platform now runs coding sessions asynchronously in the cloud, letting developers teleport a local CLI session to remote infrastructure while a 128B open-weight model works on PRs in isolated sandboxes. Paired with the Medium 3.5 release, it's Mistral's most direct challenge yet to Devin and GitHub Copilot Workspace.
Original source## What Launched
Mistral today released Mistral Medium 3.5 — a 128B dense model with open weights under a modified MIT license — alongside Vibe, a remote agent platform that runs coding sessions in the cloud.
The headline Vibe feature: developers start a task locally via CLI or Le Chat, then "teleport" the session to the cloud. The session runs asynchronously in an isolated sandbox, preserving local history and approval state. When the task completes, the agent can automatically open a GitHub pull request. Developers can monitor and approve sensitive steps without blocking on the execution.
## The Numbers
Medium 3.5 posts 77.6% on SWE-Bench Verified — a competitive but not table-topping score. API pricing is $1.50/M input and $7.50/M output. The open weights are available through NVIDIA NIM and GPU endpoints.
## Le Chat Work Mode
Beyond coding, Mistral also shipped a Work Mode in Le Chat that handles multi-step agentic tasks: cross-tool workflows spanning email, calendar, and messaging; research synthesis from internal and external sources; and inbox triage with drafted replies. All actions are transparent, and the agent requests explicit approval before doing anything sensitive.
## Who This Is For
Vibe targets developers who want Devin-style async coding but don't want full vendor lock-in — since the underlying model weights are open, teams can self-host the inference and build their own agent runner. The $1.50/M input pricing makes API-based use accessible for smaller teams.
## The Competitive Picture
This launch directly challenges Devin (Cognition), GitHub Copilot Workspace, and Google's Gemini remote coding agents. Mistral's differentiator is the open-weight backbone: if you outgrow the managed service, you can take the model weights and run them yourself. That's a meaningful guarantee that none of the closed-source competitors can offer.
Panel Takes
The Builder
Developer Perspective
“The 'teleport local session to cloud' UX is clever — you get the context of a local dev environment with the compute of a remote agent. And open weights as the escape hatch from vendor lock-in is a genuinely compelling pitch.”
The Skeptic
Reality Check
“77.6% on SWE-Bench is strong but still behind the top closed models. 'Public preview' means rough edges. And the modified MIT license needs a careful read — 'open weight' doesn't always mean unrestricted commercial use.”
The Futurist
Big Picture
“Open-weight remote agents are the architecture that survives long-term. Mistral is building the model equivalent of Linux: open enough to audit and self-host, competitive enough to use in production. That's a durable market position.”