Meta has released Muse Wonder and Muse Wonder Full, its latest frontier model series, positioning them as the company's most capable AI to date and framing the launch around the concept of "personal superintelligence." The models feature two marquee capabilities — Active Memory (persistent context about users across conversations) and Active Focus (sustained, multi-step task execution) — and claim state-of-the-art performance on agentic benchmarks spanning law, finance, health, and coding. The technical claims are substantial but selectively presented. Meta cites results on Harvey's Legal Agent Benchmark, FocusBench, and SWE-Marathon for long-running software projects, along with a novel out-of-distribution generalization analysis showing Muse Wonder Full outperforms its predecessor Muse Spark 1.3 on unfamiliar tasks at equivalent in-distribution performance. The models were trained at Meta's gigawatt-scale Prometheus data center, and Meta explicitly credits stronger base models — not just post-training optimization — for the generalization gains. The "personal superintelligence" framing is doing heavy rhetorical work. What Meta describes is a capable agent with browser access, calendar integration, and document handling — useful, but definitionally not superintelligence by any established meaning. The term appears designed to set narrative expectations for investors and the public, creating a goalpost that Meta can claim to be approaching regardless of what the models actually achieve. Fourteen months from "rebuilding our stack from first principles" to "personal superintelligence" is a timeline that rewards marketing velocity over scientific precision. The distribution strategy reveals the real architecture of value capture. Muse Wonder is available through Meta's Model API and Muse Code, with Muse Wonder Full gated behind a Muse Code subscription. The models will "soon power experiences in Muse" with subscribers getting first access. This is a classic platform flywheel: the most capable model requires the deepest platform integration, which generates the most user data (Active Memory explicitly collects personal context), which improves the model, which deepens lock-in. The safety section deserves scrutiny. Meta claims Muse Wonder is "our most aligned model to date" with "strongest safeguards" against CBRN and cyber misuse, while "preserving access to frontier cyber capabilities for defensive use." The full safety report is promised "soon" but was not released alongside the model — a pattern that allows the capability announcement to circulate without the safety evaluation attached. The tension between a model designed to handle private documents, messages, and calendars and one that must "exercise discretion with private information" is acknowledged but not resolved with external verification. The out-of-distribution generalization finding is the most technically interesting claim. If validated independently, the result that stronger base models yield post-training gains that transfer more effectively to unfamiliar tasks would have real implications for how the industry allocates compute between pretraining and post-training. But this is self-reported on internal benchmarks, and the specific tasks and metrics underlying the in-distribution and out-of-distribution composites are not detailed in the announcement. Meta is executing a coherent strategy: build the most capable agent, make it dependent on personal data ingestion, gate full capability behind subscriptions, and use "personal superintelligence" framing to justify the data access as aspiration rather than extraction. The technology may be genuinely frontier. The business model is as old as platform capitalism.
