Meta Launches Muse Wonder and Muse Wonder Full, Its First Frontier AI Models
TL;DR
- Meta's Superintelligence Lab has launched Muse Wonder and Muse Wonder Full, its first natively frontier models built for reasoning, coding, and multimodal agency.
- Muse Wonder Full tops key benchmarks in math, science, and agentic coding, putting Meta in direct contention with OpenAI's GPT-5, Google's Gemini 3, and Anthropic's Claude.
- The launch signals Meta's shift from open-weight Llama to closed frontier systems, with API access, Meta app integration, and massive compute scale-up to follow.
What Just Happened
Meta is officially in the frontier race. On Friday, Meta's Superintelligence Lab, the elite unit formed earlier this year under Alexandr Wang, unveiled its first two state-of-the-art models: Muse Wonder and Muse Wonder Full.
The announcement marks a sharp break from Meta's Llama era. Unlike Llama 3 and Llama 4, which were positioned as open-weight community models, Muse is a closed, highly capable frontier family trained on Meta's sprawling new compute clusters. In a memo shared internally and in a public blog post, Meta CEO Mark Zuckerberg called it "the first step toward personal superintelligence for everyone."
Early access is rolling out now to select developers and partners, with API access via Meta's developer platform and integration into Meta AI, Instagram, WhatsApp, and Ray-Ban Meta glasses expected in the coming weeks.
Meet Muse Wonder and Muse Wonder Full
Muse comes in two flavors. Muse Wonder is positioned as the fast, efficient everyday model — designed for low latency, high-volume chat, voice, and on-device and edge use cases. Meta says it delivers near-frontier performance at a fraction of the inference cost.
Muse Wonder Full is the heavyweight. It's a large-scale reasoning model with deep chain-of-thought capabilities, native multimodality across text, image, video, and audio, and a massive context window reportedly up to 1 million tokens. It's built for autonomous agents, long-horizon research, advanced coding, and scientific problem-solving.
Both models were trained with new post-training techniques focused on self-verification, tool use, and open-ended reasoning, areas where Meta has been aggressively hiring researchers from OpenAI, Anthropic, Google DeepMind, and Scale AI.
Benchmarks and Performance
Meta isn't being shy about performance claims. According to its technical report, Muse Wonder Full achieves state-of-the-art or matching results on several closely watched benchmarks.
On competitive coding and software engineering, Meta says Wonder Full scores at the top on SWE-Bench Verified and Terminal-Bench 2.0, outperforming OpenAI's GPT-5-high, Anthropic's Claude Opus 4.1, and Google's Gemini 3 Pro in internal evaluations. On math and science reasoning, it posts leading scores on GPQA Diamond, MATH-500, and ARC-AGI-2.
Meta also highlighted agentic and multimodal wins. Wonder Full reportedly leads on BrowseComp, CharXiv-Reasoning, and a new internal benchmark called MusePlan for multi-step web tasks involving booking, shopping, and research. The company also demoed the model autonomously building full-stack apps, analyzing hour-long videos, and controlling a web browser and Meta's smart glasses in real time.
Independent verification is still pending, and as always with vendor-reported benchmarks, outside testing via LMArena and third-party labs will be critical in the coming days.
Why This Changes the AI Race
For two years, the frontier race has been a three-way fight between OpenAI, Google, and Anthropic. Meta, despite its massive investment in Llama and open source, was widely seen as a generation behind.
Muse changes that narrative overnight. If the benchmark gains hold, Meta now has a credible alternative for enterprise developers who previously had to choose between GPT, Claude, and Gemini.
The timing is aggressive. OpenAI is preparing its next GPT release, Google is embedding Gemini 3 deeply into Search and Android, and Anthropic just closed massive enterprise deals for Claude in coding and finance. Meta is betting it can leapfrog with superior distribution — 3 billion users across Facebook, Instagram, and WhatsApp — plus unmatched compute spend, with the company pledging hundreds of billions toward data centers through 2028.
What’s Open vs. Closed
Perhaps the biggest controversy is Meta's pivot away from open source. Zuckerberg long championed open-weight Llama as the path to democratize AI. Muse Wonder and Muse Wonder Full will launch as proprietary API-only models, with no weights release announced.
Meta says smaller Muse derivatives and continued Llama updates will still serve the open community, but researchers and startups are already questioning whether Meta is abandoning its open-source promise to chase frontier prestige and revenue.
Regulators will also be watching. With antitrust scrutiny around Meta's AI talent acquisitions and data use for training, a powerful closed model embedded across its social apps is likely to draw fresh questions in Washington and Brussels.
What’s Next
Meta says this is just the beginning. The Superintelligence Lab is already training larger successors to Wonder Full, scaling test-time compute, and building personal agent features that remember user context across Meta AI, Quest, and wearables.
Developers can join a waitlist for the Muse API starting today, with general availability and pricing expected later this month. A full system card, safety evaluation, and red-teaming results are promised alongside the wider release.
The message from Menlo Park is clear: Meta no longer wants to power other people's AI revolution. With Muse Wonder and Muse Wonder Full, it wants to lead it.
Get All The Latest Updates Delivered Straight To Your Inbox For Free!