Meta launches Muse Wonder and Muse Wonder Full models
Meta shipped its biggest AI release of the day, while Microsoft's president argued the technology he sells is regulated more weakly than grocery dairy — and a Microsoft Research paper explained why long-running agents keep settling for worse answers.
Meta released Muse Wonder and Muse Wonder Full, calling them its most capable models yet and the next step on its scaling ladder. Both were trained at Meta's gigawatt-scale Prometheus data center, take text, image, video and audio input across a 1M-token context window, and ship with two new headline capabilities: Active Memory, which carries context about the user across conversations, and Active Focus, which keeps the model working through a task instead of giving up at the first blocker. Availability is deliberately narrow — Muse Wonder is live today through the Meta Model API and Muse Code, while Muse Wonder Full is reserved for Muse Code subscribers, and Meta has published no open weights, citing a lifecycle-based weight security program. The Verge reports Meta claims the models can stand next to GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1, though the proof is internal testing, and the headline benchmark numbers appear only in charts rather than published figures.
The release matters because it marks a real shift in Meta's posture: the earlier Muse Glimmer was an open-weight 30B release, while Wonder is API-gated and closed. That makes Meta's Superintelligence Labs look more like its frontier rivals and less like an open-source benefactor — expect the "personal superintelligence" framing to do a lot of work until independent evaluators get access. We've tracked the Muse line closely since the agent launch — Meta ships Muse: a consumer agent inside a sealed VM.
Microsoft president Brad Smith says AI is less regulated than a carton of milk — and wants kill switches written into law. "Right now, your washing machine is subject to more safety rules than AI. Every carton of milk that you see in the grocery store is subject to safety rules and inspection," Smith told the BBC, calling AI a "tool and a weapon" that needs urgent regulation. His asks are concrete: mandatory "emergency brakes" to pause AI systems, a legal obligation to notify victims when a rogue agent breaches their system, and independent pre-release evaluation by government testers. The comments put Smith directly at odds with the White House's preference for self-regulation, and they land days after CEO Satya Nadella separately called for an "emergency brake" on superintelligence — two Microsoft leaders arguing, in public, that the industry they lead cannot be trusted to police itself.
Clearing an AI agent's chat history makes it find better solutions, a Microsoft Research paper finds. Eight runs of the same research agent on the same task finished anywhere from 68.3 to 100.0, and the gaps survived extra compute — the agents were stuck refining answers inside what the authors call "idea basins," steered there by their own accumulated conversation history. Their fix, Fork-and-Flush, splits the work into parallel branches that keep the files and code but throw away the chat context; across 13 long-running tasks it hit a normalized score of 0.78 against 0.47 for plain single runs, a 66 percent gain on the identical compute budget, with simply clearing the context worth 22 percent on its own. The implication is uncomfortable for anyone building on long agent sessions: more context is not automatically more progress, and forgetting may be a feature.
What to watch: Meta's full safety report for the Wonder family, promised "soon," and whether independent evaluators can reproduce the self-reported benchmark claims.
Would you trust a frontier model that publishes its scores only in charts? Tell us in the comments.
