Google Cloud Adds Meta Muse Ember, Spark 1.3 and Wonder

Three Meta Muse models — Ember, Spark 1.3 and Wonder — can now be called as managed APIs on Google Cloud. Mark Zuckerberg wrote on the evening of 9 November: “Muse Ember, Muse Spark 1.3, and Muse Wonder are now available on Google Cloud.”

Google’s documentation for Gemini Enterprise Agent Platform lists the three as partner models, served on a global, OpenAI-compatible Responses API. The same page is explicit that Meta models are “not a Google product”, and that their availability is subject to the terms for “Separate Offerings” in the AI/ML Services section of Google’s Service Specific Terms, plus separate terms on the relevant model card.

This is a second distribution path, not a new model family. Earlier the same day, Meta had already released Muse Ember on its own Muse Code and Meta Model API. The Google Cloud route is a different door, with Google’s own descriptions of what each model is for, and with a stateless API that does not support stored responses.

Google Cloud serves the Muse models on a stateless Responses API

The docs say the Muse models are available as managed APIs on Agent Platform, that they can be found and enabled from Model Garden, and that responses can be streamed with server-sent events. They are served through the “stateless OpenAI-compatible Responses API” on the global endpoint. Stateful response storage (store: true) is unsupported. Callers pass a bare model ID — muse-wonder-1.0, muse-ember-1.0 or muse-spark-1.3 — in the model field. The documentation page is stamped “Last updated 2026-11-10 UTC.”

Sundar Pichai replied the same evening: “Congrats and welcome to the model garden, excited to serve our Google Cloud customers!” The welcome is real. It does not make the models a Google product, and it does not add pricing or quota figures, which the partner-models page also does not spell out.

Google describes Wonder, Ember and Spark 1.3 as a complementary set

Google’s page characterises Muse Wonder as “a high-capability reasoning model built for complex software engineering and long-horizon agentic work.” In multi-step coding, autonomous agent and computer-use workflows, it says, Wonder “plans tasks, delegates to sub-agents, calls tools in parallel, and recovers from errors.”

Muse Ember is described as “an efficient reasoning model built for high-volume agent tasks, including tool calls, coding subtasks, routing, extraction, and long-document processing.” Google tells customers to “pair it with Muse Wonder or Muse Spark 1.3 for multi-step planning and reasoning.”

Muse Spark 1.3 is “a reasoning model trained for agentic workflows and competitive coding.” Google says it supports built-in tool calling with Model Context Protocol (MCP) support and a 1-million-token context window for multi-step tasks.

Those are Google’s product-page descriptions. They sit beside Meta’s own launch language: Ember as the volume tier on 9 November, and Wonder and Wonder Full as Meta’s “most capable models to date” when they launched in October. The Google Cloud page does not republish Meta’s Ember score tables or prices, and it should not be read as if it did.

A second door is not a new Muse family

The timing matters because the two 9 November announcements can blur together. TSN’s Ember post went out at 18:43 GMT Europe/London, covering Meta’s own API SKU, proposed pricing and vendor benchmarks. Zuckerberg’s Google Cloud post followed at 21:49 GMT the same evening. One is a model launch on Meta’s rails. The other is access, under a Google-managed API, with a contractual wrapper that keeps the models outside Google’s own product line.

That distinction is the practical one for teams already on Google Cloud. They can now point an OpenAI-compatible client at the documented global endpoint for Meta models. They should also expect the documented limit: this Responses API is stateless, so the store: true pattern some OpenAI-style workflows rely on is unsupported. That is a platform constraint, not a minor footnote, if a design assumes stored response objects.

It is also a different surface from Muse Spaces, Meta’s in-app customisation feature, and from Muse Spark Open, the weights Meta released for local runs that Apple demonstrated across Mac Studios. Those are Muse stories. They are not this endpoint.

The next thing worth checking is whether Google publishes pricing and quotas for muse-ember-1.0, muse-spark-1.3 and muse-wonder-1.0 on Agent Platform. The partner-models page last updated on 10 November does not. Until it does, the fact in hand is narrower, and still useful: the three models are on Google Cloud as managed APIs, under separate Meta terms.

Sources

1. Google Cloud documentation, “Muse models from Meta,” last updated 10 November 2026. https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/meta
2. Mark Zuckerberg on X, 9 November 2026, 21:49 GMT Europe/London. https://x.com/finkd/status/2119910354369257529
3. Sundar Pichai on X, 9 November 2026, 22:41 GMT Europe/London. https://x.com/sundarpichai/status/2119923365926551765

Related stories

Meta Releases Muse Ember, Its Smallest Agent Model is the same-day launch of Ember on Meta’s own API, the post that this Google Cloud distribution story follows.

Also relevant:

• Meta Introduces Muse Spaces for Custom In-App Experiences
• Meta Launches Muse Wonder and Muse Wonder Full, Its “Most Capable Models to Date”, Built for AI Agents