Meta's Muse Wonder Advances in Real-World Agent Evaluation, Outperforms Spark 1.3

Meta has released Muse Wonder, a frontier-class language model designed for extended, multi-step agentic tasks, with early evaluations on the Agent Arena real-world benchmark showing measurable performance gains over its predecessor Muse Spark 1.3. Announced in October 2026, Muse Wonder processes 1 million token contexts and supports multimodal inputs—text, image, speech, and video—while outputting text at 131 tokens per second. The model includes two key features for agentic workflows: Active Memory, which recalls context from prior conversations to maintain continuity across extended sessions, and Active Focus, which enables error recovery, task persistence, and confirmation before critical actions.