Meta Launches Muse Wonder as Zuckerberg and Wang Detail $100B AI Bet

Muse Wonder Launches as Meta's New Frontier Model

Meta has launched Muse Wonder, a new class of frontier model designed to serve as a personal and professional agent. The model comes in two versions: Muse Wonder and Muse Wonder Full, the latter being a much higher-end variant. According to Alexandr Wang, Chief AI Officer at Meta Superintelligence Labs, the models show state-of-the-art performance across evaluations on personal and professional agent tasks.

Key capabilities built into Muse Wonder include active memory, multimodal perception, and frontier coding. Wang said the team invested disproportionately in memory compared to other efforts, with internal evals called Memory Bench demonstrating the model's capabilities in this area.

The Model Is the Product

Zuckerberg and Wang emphasized that Muse Wonder reflects Meta's internal philosophy that the model is the product. Much of the product's personality, behavior, and capabilities are trained directly into the model rather than built as separate software layers. Wang noted that intentional decisions around how the model behaves in specific scenarios, its personality, and its handling of tool-use cases are core differentiators that are not immediately visible to users.

The Muse product notably has no model selector or effort selector, a deliberate choice to keep the consumer experience simple. Zuckerberg contrasted this with other AI products that use confusing model naming conventions, which Wang said are designed by highly technical teams without a randomly sampled consumer in their mental model.

Infrastructure Scale and Talent Reboot

Meta is scaling its AI infrastructure past five gigawatts. Zuckerberg revealed that the company committed over $100 billion in capital expenditure before knowing whether the model would work, describing a moment where he asked Wang directly if the investment would pay off.

The launch follows a significant internal restructuring. Zuckerberg said Meta was not happy with the trajectory of Llama 4 and rebuilt the team over the past year and a half, a process he described as a complete talent reboot. The team built up different tiers of models methodically before reaching the frontier with Muse Wonder, and Zuckerberg said the time between each subsequently bigger class of model will now start to compress.

Sources: Cheeky Pint · Techmeme