The new collaboration between Mark Zuckerberg and Sundar Pichai places Meta’s proprietary Muse models side by side with models from Anthropic, xAI, and DeepSeek within the Google Cloud model garden.
Zuckerberg’s announcement and Pichai’s welcome of Muse models to Google Cloud
Technical specifications for the preview via the Gemini API for enterprise agents
Task division between Muse Wonder, Spark 1.3, and high-density Ember
TerminalBench 4.0 benchmark results and implications of the shared cloud distribution
Frequently asked questions
Zuckerberg’s announcement and Pichai’s welcome of Muse models to Google Cloud
Meta has made three of its artificial intelligence models from the Muse family available to Google Cloud platform customers. As CEO Mark Zuckerberg wrote in a post on X, reported by The Economic Times in its coverage of the announcement: “Muse Ember, Muse Spark 1.3, and Muse Wonder are now available on Google Cloud.”
Google CEO Sundar Pichai responded by welcoming the move: “Congrats and welcome to the model garden, excited to serve Google Cloud customers!” This deal places Meta’s proprietary models within a direct competitor’s cloud marketplace alongside Anthropic, xAI, and DeepSeek models, just one day after the launch of the smaller and more cost-effective Ember model.
Technical specifications for the preview via the Gemini API for enterprise agents
Release notes published by Google on November 9 indicate that Muse Wonder 1.0, Muse Ember 1.0, and Muse Spark 1.3 are available in preview via the global endpoint through an OpenAI-compatible responses API. All three models support reasoning, function calling, structured outputs, and a one-million-token context window.
The Wonder and Spark 1.3 models accept text, image, and PDF document inputs, while Ember on Google Cloud accepts text and images. Google developer documentation notes that the models operate as managed APIs on the Gemini for enterprise agents platform, with the disclaimer that they are not Google products and are subject to separate terms. Google Cloud Tech’s account confirmed they are delivered “with enterprise-grade security and serverless scaling.”
Task division between Muse Wonder, Spark 1.3, and high-density Ember
Google describes Wonder as a high-capability reasoning model tailored for complex software engineering and long-horizon agentic work capable of delegating tasks to sub-agents and recovering from errors. Spark 1.3 is described as a reasoning model trained for agent workflows and competitive programming with built-in Model Context Protocol support.
Ember targets high-density workloads repeated by agents throughout the day, priced at $0.10 per million input tokens and $0.50 per million output tokens. Zuckerberg wrote that it “delivers higher intelligence per dollar compared to others in its class,” and Alexander Wang, head of Meta Superintelligence Labs, confirmed it is the first model released by the company after being trained using Muse Waterfall.
TerminalBench 4.0 benchmark results and implications of the shared cloud distribution
Alongside Ember matching Claude Haiku 5.5’s score on the synthetic analysis index at roughly one-third of the cost, Meta announced that the model scored 40.9% on the TerminalBench 4.0 benchmark, outperforming Haiku at 39.2% and OpenAI’s GPT-6 Luna at 16.4%, though noting these figures are released by Meta and have not yet undergone independent verification.
Making the Muse family available on Google Cloud—alongside the expansion of the Muse smartwatch agent—reflects Meta’s push to commercially market its closed models through major cloud platforms for direct enterprise customer access.
The addition of the three Muse models to the Google Cloud model garden signals a striking strategic shift in competitive dynamics among tech giants. Major corporations are moving past direct competition sensitivities to maximize cloud infrastructure returns and broaden the distribution of their proprietary models among enterprise customers, while independent performance and cost benchmarks remain the decisive standard for developer choices. This managed availability will allow organizations to directly compare the performance and cost of Muse models against Gemini and Claude models within a single, unified cloud environment.
Frequently asked questions
Q: What are the three Meta models that have become available via Google Cloud?
A: They include Muse Wonder 1.0, Muse Ember 1.0, and Muse Spark 1.3 in preview via the model garden and the Gemini for enterprise agents platform.
Q: What technical features are shared among the three Muse models on Google Cloud?
A: All of them support reasoning, function calling, structured outputs, and a one-million-token context window via an OpenAI-compatible API, with Wonder and Spark also accepting text, images, and documents.
Q: How do Muse Wonder and Muse Spark 1.3 differ in enterprise tasks?
A: Wonder specializes in complex software engineering, sub-agent delegation, and error recovery, while Spark 1.3 focuses on competitive programming and Model Context Protocol support.
Q: What score did the Muse Ember model achieve on TerminalBench 4.0 according to Meta’s data?
A: It scored 40.9% compared to 39.2% for Claude Haiku 5.5 and 16.4% for GPT-6 Luna, based on results announced by Meta.