The Muse Wonder models cap off 14 months of restructuring and intensive investment within Meta to establish its presence in the front lines of superintelligence computing.
Launch of Muse Wonder and restructuring of Meta’s AI division
Standard benchmark results and outperforming competitors
Active memory and enhanced focus for solving complex problems
Advanced safety controls and open-source policy
Frequently asked questions
Launch of Muse Wonder and restructuring of Meta’s AI division
Meta Platforms officially announced on Monday, October 12, 2026, the rollout of its new smart model family under the name “Muse Wonder” and the ultra-capable version “Muse Wonder Full,” describing them as the most advanced and capable systems in the company’s history, capable of competing with the top technological solutions developed by OpenAI and Anthropic.
Bloomberg reported that this launch marks the most prominent and advanced achievement for the company since CEO Mark Zuckerberg began a comprehensive and bold rebuilding of the artificial intelligence division through an intensive recruitment and talent acquisition campaign lasting about 14 consecutive months. This rollout represents a critical test of the viability of those massive multibillion-dollar expenditures, with the base Muse Wonder model made available to developers via the API and the Muse Code programming platform, while the Wonder Full model was reserved for exclusive subscribers, prompting Meta’s stock to rise immediately following the announcement.
Standard benchmark results and outperforming competitors
Meta published performance benchmarks and technical evaluations for the models, which carried the internal codename “Watermelon,” showing them fiercely competing with or outperforming the latest products on the global market. Wired magazine reported that on the independent artificial intelligence analysis index, the larger version of Muse Wonder ranked second directly behind Anthropic’s “Claude Opus 5.5,” while Wonder clearly outperformed OpenAI’s “GPT-6 Astra” in rigorous specialized tests involving the review of legal documents and the development of complex code.
Alexander Wang, head of Meta Superintelligence Labs, told Axios: “These are true frontier models competing with the finest intelligent systems in the entire world, representing the vanguard of technology and a state of supreme mastery in executing independent personal and professional agentic tasks.”
Active memory and enhanced focus for solving complex problems
Through this family, Meta introduced two technical innovations that change the rules of smart interaction: the first innovation is “active memory,” which retains user context, details, information, and prior work history across successive conversation sessions without forgetting, while the second innovation is “active focus,” which allows the AI agent to continue processing and solving challenging mathematical and programming problems without giving up, by exploring alternative solution approaches and asking the user for additional clarifications if necessary.
Meta confirmed that the training of these giant models took place entirely inside its high-capacity “Prometheus” data center operating on gigawatt-scale power, granting them massive capacity to process millions of data points and real-time instructions with exceptional accuracy and stability.
Advanced safety controls and open-source policy
Meta confirmed that its new models were trained to strictly refrain from leaking any private information or personal data without explicit authorization, and to firmly reject any instructions that could lead to material or moral harm, in accordance with the internally adopted “Superintelligence Expansion Framework” standards, with a pledge to publish the comprehensive technical report for the model’s safety card.
This step aligned with Zuckerberg’s statements to the Financial Times in which he confirmed Meta’s ascent to the ranks of pioneers and rejected calls for technical deceleration. Regarding the issue of making the Wonder family available as open-source software, Alexander Wang avoided making a definitive commitment at present, but emphasized the company’s continued general commitment to the open-source approach in its future projects, alongside the ongoing provision of Muse personal assistant subscriptions at $20 and $100 per month tiers.
Frequently asked questions
Question: What are the new Muse Wonder models launched by Meta?
Answer: They are Meta’s latest flagship artificial intelligence models, including the Wonder and Wonder Full versions directed at developers and programmers for complex and independent tasks.
Question: How did the new models perform compared to products from OpenAI and Anthropic?
Answer: The models achieved very advanced rankings, placing directly behind Claude Opus 5.5 and outperforming GPT-6 Astra in legal contract drafting and coding tests.
Question: What are the active memory and active focus features?
Answer: Active memory retains user context and information across extended conversations, and active focus allows the model to continuously attempt to solve challenging problems and explore solutions.