Mistral Large 4 open weights: Oct 31 ETA and what to do now

Mistral Large 4 is not downloadable yet. Mistral's own Hugging Face page, read on 2026-10-08, shows it as an upcoming release with the current ETA "October 31, 2026". Sume's agent model catalog does not list it, and the Agent Completions model field accepts only sume-agent, so nothing in a Sume integration changes on that date.

What the model page says

The vendor page describes a 1 trillion-parameter, natively multimodal model with 49 billion active parameters per token (52 billion including embeddings and output layers). It says open weights drop at the end of the month. It does not state a context window or a license in the text I could read, so this post makes no claim about either.

What it means for a Sume video agent

The Agent Completions docs say the model field accepts only sume-agent, and a different value is a 400 invalid_request. The agent behind that id is the same runtime as the Agents chat UI, with a sandbox, tools, the MCP bridge, and media generation. You do not choose Mistral, Claude, or GLM per call there.

Separately, the model catalog in Sume's repository, read on 2026-10-08, has no Mistral Large 4 row. So Sume does not offer it as an agent model today. When open weights land, they will matter to teams that self-host a model for their own text work, not to calls made to the Sume API.

Do now, do later

The points that matter here, in the order you will hit them:

• Do not plan a launch around an ETA; it is a current estimate on a model page.
• If you self-host, size the hardware from the final model card, not from the parameter count alone.
• Keep Sume integration code on model: "sume-agent" or omit the field.
• Re-read the Sume model picker and the docs after the release, rather than assuming a name appears.

A rule for any new model name

Treat "does Sume offer model X" as a question for the catalog, not for the announcement. Announcements tell you a model exists. Only the product's own list tells you that you can call it.

Weights are not a pipeline

Open weights give you a text and image model that you must serve yourself. A Sume agent run is more than a model. The docs describe a sandbox, tools, an MCP bridge, and media generation, and the receipt returns durable media URLs and a usage record. Even if Mistral Large 4 behaves well when released, reproducing that pipeline around it is a separate project. That is the main reason a new open model changes your hardware plans before it changes how you call a hosted agent.

A final practical point is the license. The page text I could read does not state one, so do not assume you may use the weights commercially until the release includes a license file you have read. Treat the ETA as a date to re-check, not a date to ship against.