Microsoft Puts Meta's Muse Models on Foundry, No Date or Price
Microsoft will add Meta's Muse Wonder, Spark 1.3 and Ember to Foundry, but gave no date or pricing. The move puts rival frontier models on one cloud platform alongside GPT, Claude and MAI.
By Olivia Hart
4 min read

What's News
- Microsoft announced on November 17, 2026 that Meta's Muse Wonder, Muse Spark 1.3 and Muse Ember are 'coming soon' to Microsoft Foundry, with no date or pricing disclosed.
- The Foundry catalog already lists more than 11,000 models from Microsoft, OpenAI, Anthropic, Meta, Mistral and open-source providers.
- Microsoft claims MAI-Code-2 delivers 98% of GPT-6.1 Sol's performance at 40% of the cost—based solely on internal benchmarks.
- The Maia 300 accelerator is claimed to deliver 2.5x high-bandwidth memory performance over its predecessor; Cobalt 200 expands to 23 Azure regions by year-end.
- MAI-Thinking-1, a 35B-parameter sparse mixture-of-experts reasoning model, launched in preview at Ignite.
Microsoft announced on November 17, 2026, at its Ignite conference in San Francisco that three of Meta's newest Muse AI models—Muse Wonder, Muse Spark 1.3, and Muse Ember—are "coming soon" to Microsoft Foundry. No release date, no pricing, and no regional availability details accompanied the announcement. The models appear in the Foundry catalog with only that two-word promise.
The move places models from two frontier AI competitors side by side on the same cloud service, alongside OpenAI's GPT series, Anthropic's Claude, and Microsoft's own MAI family. Foundry's catalog already lists more than 11,000 models from providers including Microsoft, OpenAI, Anthropic, Meta, Mistral, and the open-source community. In plain terms: Microsoft is selling rivals' products next to its own.
What are the Muse models supposed to do?
Microsoft's blog describes the trio as a tiered set: one tier for frontier long-horizon planning and complex coding, one balanced tier, and one for fast, cost-efficient execution at scale. The post says the models give teams "frontier, balanced and high-efficiency options to match each step of an agentic workflow."
That is positioning language, not shipping documentation. Until the models arrive and independent benchmarks exist, the tiering is a promise about workflow coverage, not a tested capability.
What did the CEOs say?
Satya Nadella posted on X that Microsoft is "giving developers and enterprises even more choice and flexibility." Mark Zuckerberg separately confirmed the three Muse models were "coming soon" to Foundry. The Foundry blog post states: "Their expertise should strengthen their business, not deepen their dependence on any one AI lab."
That anti-lock-in framing is the strategic core of the deal. It also sits awkwardly beside Microsoft's heavy investment in OpenAI and its parallel buildout of the MAI family—Microsoft is funding one frontier lab, building another, and hosting all of them.
What else launched at Ignite?
The Meta announcement was one piece of a much larger rollout. Microsoft introduced or updated its entire MAI family in a single day:
- MAI-Code-2 and MAI-Code-2-Flash, for agentic coding and high-volume everyday coding respectively
- Microsoft-Decision-1.1, a Gemma-based model for routing, classification, and verification, now generally available with image input support
- MAI-Image-2.7 and MAI-Image-2.7-Flash, for creative media generation, in public preview
- MAI-Cyber-2-Flash, for vulnerability discovery and remediation, available through the Codename MDASH security harness
- MAI-Thinking-1, a 35B-parameter sparse mixture-of-experts reasoning model, in preview
Microsoft also announced partnerships to bring Baseten-optimized open models—Kimi-K3, GLM 5.3, GLM-5.3-Flash and DeepSeek V4.1-Flash-1017—as a Microsoft first-party service. Video models from fal—Kling v3 Pro, MiniMax H3 Max, MiniMax H3 Max Turbo, LTX-2.5 Pro and LTX-2.5 Fast—will come to Foundry Models through a partnership.
Can the benchmark claims be trusted?
Microsoft says MAI-Code-2 delivers 98% of GPT-6.1 Sol's performance at 40% of the cost, and 96% of Claude Opus 5.5's performance at 5% of the cost, averaged across four coding benchmarks. For MAI-Cyber-2-Flash, the company cites its "Cyvergime E2E" evaluation showing less than 2% of GPT-6 Astra's cost per task at equivalent quality.
Every one of those figures comes from Microsoft's own internal benchmarks. As the InfoWorld analysis puts it, "treat vendor-run benchmarks as a starting hypothesis and test the model on your own workloads." Several MAI models also remain in preview, where support guarantees and expectations of change differ from generally available releases.
What hardware sits underneath?
The model push rides on Microsoft's custom silicon. The Maia 300 inference accelerator, described as a full platform spanning rack, server and networking, is said to deliver 2.5x high-bandwidth memory performance over its predecessor. The Cobalt 200 VM, Microsoft's own Arm-based CPU, will expand to 23 Azure regions by year-end, with up to 50% better performance than Cobalt 100 on some AI workloads. Managed Compute, a dedicated infrastructure option for running open and custom models on Foundry, reached general availability at Ignite.
So what?
Foundry now positions itself as a neutral layer: Microsoft models, partner models, and open models all deployable through one control plane, with governance tied to Entra, Defender and Purview. What remains unknown is when the Muse models will actually ship, what they will cost, and how they perform on independent evaluations. Until those questions have answers, the announcement signals intent—Microsoft wants Foundry to be the default marketplace for enterprise AI, and Meta wants distribution—rather than a product developers can run today.
Source: epyhia.com
More from Olivia Hart
Show full bio
Staff writer covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.
2,301 articles