Meta has announced its new AI models, Muse Wonder and Muse Wonder Full. The announcement centers on two models. The series, which follows a restructuring that began about 14 months ago, is built on Muse Spark and was trained at the Prometheus data center. The company describes the models as its most capable to date.
The models’ “active memory” feature stores important information from different conversations and reuses it when needed. The features also help sustain tasks. “Active focus” enables the models to overcome obstacles, change their approach, and ask the user for clarification when needed. Use cases cover legal documents, financial information, and health content. Software development is also among these areas. Examples include creating a running plan, a wedding seating arrangement, and developing a playable game.
Both models process text, image, audio, and video input and generate text responses. The context capacity is set at 1 million tokens.
The SWE-Marathon results ranked as follows. Wonder Full scored 57.5 points. Wonder reached 53.1 points. Claude Opus 5.5 scored 46.3 points, while GPT-6 Astra remained at 42.5 points.
Meta added a note about its testing method. The company says it tested the models with different coding tools. In Artificial Analysis’s Intelligence Index, Wonder Full scored 56 points.
Access varies by model. Muse Wonder is available through the Meta Model API and Muse Code. Wonder Full, for now, is limited to Muse Code subscribers.
Background
Meta is not a new name in the FikirPilot archive: over the last 90 days, we have published 45 articles mentioning this name; the most recent is dated October 8, 2026.