FlowtandBriefAI · Software · ResearchMeet Flowtand

Mistral moves Vibe coding agents to the cloud on a new 128B model

Announcement

Mistral AI

Browse Agents topic page →

Mistral says its Vibe coding agents can now run asynchronously in the cloud, launched from the Vibe CLI or Le Chat, with a new Work mode for multi-step tasks. Both are powered by Mistral Medium 3.5, a 128B dense model released as open weights, which Mistral reports scoring 77.6% on SWE-Bench Verified. The company prices API use at $1.5 per million input tokens and $7.5 per million output tokens.

What Mistral announced

Mistral says coding agents in its Vibe product have moved from the laptop to the cloud. Sessions can be launched from the Vibe CLI or inside Le Chat, several can run in parallel, and an active local CLI session can be handed off to the cloud with its history, task state and approvals intact. Each session runs in an isolated sandbox that permits broad edits and installs, and agents can open a pull request on GitHub and notify the user when finished. The underlying model is Mistral Medium 3.5, a dense 128B model with a 256k context window, released as open weights under a modified MIT license and positioned by Mistral as its new default, replacing Devstral 2. Mistral also previews a Work mode in Le Chat that chains tool calls for multi-step tasks with connectors enabled by default and approval prompts before sensitive actions. Mistral reports 77.6% on SWE-Bench Verified and 91.4 on τ³-Telecom; these are the vendor's own figures.

A trial scenario for an engineering lead

An engineering lead with a backlog of well-scoped maintenance work could try routing a single category, such as dependency upgrades or test generation, through Vibe remote agents rather than ad-hoc local runs. The premise comes from Mistral's description that Vibe fits high-volume, well-defined work where the developer's judgment is still needed at review time. A team could keep humans on pull-request review and on the approval prompts for any sensitive action, and reserve parallel sessions for tasks whose scope is already written down.

How to run a bounded pilot

Pick one task type and a fixed set of tickets, run them locally first to establish a baseline for review effort, then repeat on remote agents with the same review standard. Measure time to a reviewable pull request, how often reviewers request changes, and how many sessions stall or need manual rescue; treat those as failure checks, not just speed. Success would mean equal or better review acceptance at lower developer interruption, not merely more sessions spawned. Keep approval gates on and require human sign-off before merges, since Mistral describes oversight as retained rather than automatic.

Trade-offs and when to hold back

Remote sessions mean code, diffs and tool calls flow through a hosted runtime, so teams with strict data-residency or source-code handling rules should check plan and sandbox terms before adopting. Parallel agents multiply review load; without disciplined triage, the bottleneck moves to reviewers. Self-hosting the open-weight model on four GPUs, as Mistral suggests, is an option only where infrastructure and licensing review exist. Because Work mode and the model are previews, teams needing stability may wait. Skip it for ambiguous design work where step-by-step human judgment drives the outcome.

Context and limitations

  • Based on a short source excerpt of a vendor announcement; the full article was not analyzed.
  • Performance figures (SWE-Bench Verified, τ³-Telecom) are Mistral's own reported numbers, not independently verified here.
  • Work mode and the model are described as previews; availability, connectors and behavior may change.
  • Based on bounded source text, not independent testing. Practical suggestions are editorial proposals.