Meta says Muse Wonder beats Muse Spark 1.3's trend on unfamiliar tasks
AIatMeta · x · 2026-10-13
Meta AI's official account says that to quantify how models handle unfamiliar tasks, it compared test-time compute scaling on tasks seen during post-training against other tasks.
Both models improve similarly out of distribution, but at the same in-distribution score, Muse Wonder Full performs better on unfamiliar tasks than the Muse Spark 1.3 trend predicts. Meta attributes the upward shift to a stronger base model that carries post-training gains more effectively to unanticipated tasks.
The post also highlights two capabilities for long-running, multi-step agent work: Active Memory, which carries context across conversations and surfaces it when relevant, and Active Focus, which keeps the model working toward the goal when a task gets challenging, trying alternatives and following through to completion.
Related event: Meta Unveils Muse Wonder, Its Most Capable Frontier Models Yet(45 posts)→
More from Models
- Meta launches Muse Wonder models, tops SWE-Marathon at 57.5% — acbuller · 2026-10-13
- 60.6% of LLM theorem rewrites silently drop hidden hypotheses, study finds — MFarajtabar · 2026-10-13
- Meta's superintelligence lab launches Muse Wonder, a frontier model tuned for personal agents — alexandr_wang · 2026-10-13
- Muse Code CLI goes live with one-line curl install for Muse Wonder — alexandr_wang · 2026-10-13
- Meta's Wonder Full Debuts #5 on Vals Index, Outranking Every OpenAI Model — airesearch12 · 2026-10-13
- Mysterious new model Muse Wonder appears on LMArena in Battle and Agent modes — arena · 2026-10-13