Meta is SO BACK ♾️
Fourteen months after launching its AI overhaul, Meta is betting everything on "personal superintelligence," branded "Muse." The Muse Wonder series is the first answer to that bet: a flagship model for personal and professional scenarios, built as an Agent rather than a traditional general-purpose chat model.
Four Pillars: From "Answering Questions" to "Doing Things for You"
1. Domain Expertise: Not just common-sense Q&A, but handling specialized material with real consequences, such as leases, finances, and health. Evaluations cover law, finance, and medicine, including Harvey's legal agent benchmark (LAB), which requires models to produce legal work product from original documents.
2. Active Memory: Carries the user's key background information across conversations and proactively retrieves relevant details. This is the core dividing line between a personal Agent and a general assistant: it has to "know you."
3. Active Focus: The most noteworthy piece. It means the model can stay engaged with difficult goals: switching approaches, working around obstacles, adjusting strategy based on new information, and when necessary proactively asking the user for clarification or confirming before important actions. Two in-house benchmarks accompany it: FocusBench (focus) and SWE-Marathon (sustained execution on long-horizon software projects). "Not giving up or drifting off course during long tasks" is explicitly treated as a first-class capability that can be trained and evaluated, a direct response to the industry's pain point of Agents "quitting halfway."
4. Frontier Coding: From small apps to complex software systems, with training focused on diverse long-horizon tasks, meaning the model keeps making progress through multiple rounds of test-and-revise cycles.
The Most Informative Part: The Argument for Generalization
Meta AI responds to two contradictory industry reactions: criticism of "benchmaxxing" and praise for the "big model feel." Both actually point to the same thing: the breadth of capability that popular benchmarks fail to capture.
Meta's analysis is methodologically valuable. They plotted the performance of Muse Spark 1.3 and Muse Wonder Full, with the x-axis showing performance on in-distribution (training-target) tasks and the y-axis showing performance on unseen out-of-distribution tasks, with each point representing a different inference compute tier.
The conclusion: as test-time compute increases, OOD performance rises in step for both models; but Muse Wonder Full's OOD scores exceed what the Muse Spark 1.3 trend line predicts from its ID scores. In other words, a stronger base model means that post-training gains transfer more effectively to unfamiliar tasks.
The strategic implication of this conclusion is larger than the technical one: a personal Agent faces an unpredictable combination of goals, tools, and constraints, so generalization is the moat. Meta is effectively arguing that "building a bigger, stronger base" and "post-training fine-tuning" are not a zero-sum trade-off but a compounding relationship; the stronger the base, the greater the spillover from fine-tuning. This provides empirical support for continuing aggressive scaling.