Mistral Medium 3.5 — Mistral ships an agent harness with the model
Mistral didn't just drop a new model this week. They shipped an agent harness alongside it.
Medium 3.5 is a 128B-parameter dense model with a 256k context window, released as open weights under a modified MIT license. 77.6% on SWE-Bench Verified, 91.4 on the τ³-Telecom agentic benchmark. API pricing is $1.5 per million input tokens, $7.5 per million output—materially cheaper than Anthropic's flagship.
The bigger story is what shipped with it: Vibe Remote Agents. Cloud-based coding sessions that execute asynchronously in sandboxes, integrated with GitHub, Linear, Jira, Sentry, Slack, and Teams. Open the CLI, kick off an agent, close your laptop, come back to a PR. Session teleportation lets you start work locally and move it to the cloud mid-task. Same pattern Cognition shipped with Devin for Terminal two days ago.
The interesting bet is on distribution, not the model itself. Anthropic has Claude Code, OpenAI has Codex CLI, Cognition has Devin. Mistral was the European-sovereign-cloud, run-it-yourself option without a clear agent client. This closes the gap. For European enterprises that need both an open-weight model and a working agent stack, Mistral is now a one-stop answer—a position that maps cleanly to the procurement constraints coming out of the EU AI Act.
Memo for the foundation labs: the model layer is necessary but not sufficient. If you don't ship an agent client, you get routed around. Warp open-sourced its terminal yesterday, Cognition shipped Devin Terminal Wednesday, now Mistral has Vibe—third major bet on the agent-client layer in a single week.
Link: https://mistral.ai/news/vibe-remote-agents-mistral-medium-3-5
← Back to all articles
Medium 3.5 is a 128B-parameter dense model with a 256k context window, released as open weights under a modified MIT license. 77.6% on SWE-Bench Verified, 91.4 on the τ³-Telecom agentic benchmark. API pricing is $1.5 per million input tokens, $7.5 per million output—materially cheaper than Anthropic's flagship.
The bigger story is what shipped with it: Vibe Remote Agents. Cloud-based coding sessions that execute asynchronously in sandboxes, integrated with GitHub, Linear, Jira, Sentry, Slack, and Teams. Open the CLI, kick off an agent, close your laptop, come back to a PR. Session teleportation lets you start work locally and move it to the cloud mid-task. Same pattern Cognition shipped with Devin for Terminal two days ago.
The interesting bet is on distribution, not the model itself. Anthropic has Claude Code, OpenAI has Codex CLI, Cognition has Devin. Mistral was the European-sovereign-cloud, run-it-yourself option without a clear agent client. This closes the gap. For European enterprises that need both an open-weight model and a working agent stack, Mistral is now a one-stop answer—a position that maps cleanly to the procurement constraints coming out of the EU AI Act.
Memo for the foundation labs: the model layer is necessary but not sufficient. If you don't ship an agent client, you get routed around. Warp open-sourced its terminal yesterday, Cognition shipped Devin Terminal Wednesday, now Mistral has Vibe—third major bet on the agent-client layer in a single week.
Link: https://mistral.ai/news/vibe-remote-agents-mistral-medium-3-5
Comments