Mistral announced a bundle of updates:

  • Mistral Medium 3.5
  • remote coding agents in Mistral Vibe
  • Vibe tasks launched from Le Chat
  • Work mode in Le Chat

The model is part of the story, but I am more interested in the runtime.

Mistral describes the shift this way:

Coding agents have mostly lived on your laptop.

The rest of the announcement is about moving that work into cloud sessions that can keep running.

The model details are still useful context:

  • dense 128B model
  • 256k context window
  • open weights under a modified MIT license
  • configurable reasoning effort
  • Mistral says 77.6% on SWE-Bench Verified
  • Mistral says 91.4 on τ³-Telecom
  • $1.50 per million input tokens and $7.50 per million output tokens
  • default model in Le Chat
  • replaces Devstral 2 in Vibe CLI
  • available through Mistral, Hugging Face, and NVIDIA endpoints / NIM

But the workflow shape is the thing:

local coding agent
-> remote cloud session
-> parallel tasks
-> inspect diffs, tool calls, and progress
-> open a PR when ready

The best detail is that local CLI sessions can be “teleported” to the cloud with:

  • session history
  • task state
  • approvals

I already use this pattern with Claude Code and find it very useful. Local agents are fast and close to the developer. Remote agents are better when the task should keep going after I leave, or when I want multiple tasks running in parallel.

I do not think one replaces the other.

The primitive I keep seeing looks like this:

task + repo + tools + sandbox + permissions + long-running runtime + review surface

You might start that from a terminal, chat app, issue, incident, Slack thread, or something else. The UI matters less than the work container.

Mistral’s integration list points the same way:

  • GitHub for code and PRs
  • Linear / Jira for issues
  • Sentry for incidents
  • Slack / Teams for reporting

This is more practical to me than one magic chat box. Work already appears in those places. A useful agent runtime should be able to pick it up there and report back.

Mistral also says:

Each coding session runs in an isolated sandbox

Good. If an agent can install packages, make broad edits, run tools, and open a PR, the sandbox and approval model are not implementation details. They are the product.

Le Chat

Mistral is also turning Le Chat into a place where you can start coding sessions directly.

That fits the pattern we keep seeing: chat is becoming the place where you describe work, while the real work happens in a longer-running runtime somewhere else.

Work mode in Le Chat is broader than coding:

  • catch up across email, messages, and calendar
  • prep meeting briefs
  • research and synthesis
  • inbox triage and draft replies
  • create Jira issues
  • send summaries to Slack

I do not know yet how good this is in practice. Directionally, it lines up with where these tools seem to be going.

Octopus One-Person Band

I also tested Mistral Medium 3.5 in Msty Studio with temperature 0.7.

I used my Octopus One-Person Band prompt:

Create a single self-contained HTML file for "The Octopus One-Person Band."

It should fit in a 1200x800 screenshot.

Requirements:
- show one octopus playing exactly 4 different musical instruments at the same time
- the octopus must have 8 visible arms
- each instrument must use exactly 2 arms
- label each instrument clearly
- include a small stage, audience, and show poster
- include one small interaction using only HTML/CSS/JS
- no external libraries or images
- keep it charming, readable, and visually organized
- avoid generic concert poster styling

The result was very disappointing.

Here is the generated Octopus result.

One toy prompt does not prove much. I would not use it as a benchmark. It is just one small creative HTML test that I happen to run a lot because it exposes visual planning mistakes quickly.

It still affects how I read the launch.

Medium 3.5 may be strong where Mistral is optimizing: agentic coding, long context, long-running tool use, and enterprise workflows. My little octopus test did not make me want to use it for creative UI generation.

So my read is narrow:

local session -> cloud runtime -> reviewable PR

That bridge is the part I want to watch.