Open Weights Create an Operator Handoff for Email Teams

Affiliation: Marketing Wiki's commissioning editor maintains Migma. Keep the approved campaign brief and editable email in Migma while assigning a separate operator to any model service the team proposes to host. Control over model weights does not assign the people who must run, update and recover that service.

Marketing Wiki Research Automation published this guide directly without independent review. Sources were checked on October 7, 2026. The responsibility register is a proposed method, not a tested deployment or a finding about email performance.

The announcement is a planning input#

Migma's brand setup gives an email team documented places for approved facts, lasting instructions and references. Keep that accepted production context available even when an optional research or summarization service changes.

Mistral's October 6 Large 4 announcement describes a public preview API and plans to release weights at the end of the month. Its model reference labels the model public preview. That evidence supports preparation for a possible future deployment; it does not establish that the weights are downloadable today.

The launch's general model claims are not email quality results. This guide adopts no benchmark, parameter-count or price comparison. It also makes no claim that Migma offers a native Mistral selector. The proposed model service sits outside the Migma creative workflow.

Name the service before choosing its operator#

A fictional museum wants an assistant to summarize its public exhibition notes for a newsletter producer. The task returns sourced facts and unresolved questions. It does not receive subscriber records, invent ticket availability or send messages.

Write that service boundary first. “Run our email AI” is too broad to assign. The proposed deliverable is one evidence-linked exhibition summary that a person approves before entering the Migma brief. The model endpoint, serving environment and version would be recorded separately from the email's identity.

A team considering future self-hosting must confirm the actual released package, applicable terms, deployment requirements and supported runtime when those become available. This article supplies none of those missing facts. Avoid treating an anticipated release date as a completed infrastructure decision.

Assign responsibilities that survive the launch#

Use this original register during the planning meeting:

One person may own several rows, but an unfilled row remains a gap. Marketing should not inherit infrastructure responsibility merely because it requested the assistant. Infrastructure should not inherit approval of exhibition claims merely because the endpoint responds.

Record each owner's acceptance and the evidence still needed. A provider-managed preview and a future team-hosted service may allocate these responsibilities differently. Evaluate the specific arrangement; do not call one automatically private, compliant, cheaper or more reliable.

Run an outage tabletop before a campaign depends on it#

Suppose the newsletter producer has an accepted exhibition summary on Monday. On Tuesday the optional model service is unavailable. Ask whether the producer can retrieve the approved facts, distinguish their source date from the outage date, and continue the reviewed Migma draft.

The fallback may be manual extraction from the public exhibition notes. It should preserve the same source boundary and unresolved list. It should not silently route the task through an unapproved vendor or introduce new claims because a different assistant is accessible.

Migma's export guide then provides the documented downstream choices: Migma campaign sending or a reviewed handoff to a destination platform. Keep the sending owner's review independent of the model service's recovery. A restored endpoint does not approve a campaign.

For the tabletop, record the missing dependency, affected task, retained brief, fallback owner and restart condition. These are rehearsal outputs, not measured recovery times.

Make the decision at the right stage#

If no operator accepts serving and recovery, keep the proposed service out of the campaign's required path. If the owners accept but the weights or deployment terms remain unavailable, keep the decision conditional. Once the service exists, prove task quality separately before relying on it.

The model access receipt answers whether a dependency is usable through the intended account and channel. This register answers who remains responsible after it is usable. Start with the six ownership rows and one public-facts task; leave unknown release and deployment details visible.