Meta’s first frontier model, released 12 October 2026 by Meta Superintelligence Labs. It is larger than Muse Spark and built for long agent tasks, and it is the only Wonder model you can call through an API.
Paste a server URL, pick a model, and watch it call your tools in a real conversation. No install.
No server of your own? Leave it blank and use one of the public mock servers.
MCP Evals
Can Muse Wonder actually complete tasks with your server?
One chat shows the tool calls work. An eval writes real tasks from your tool schemas, has Muse Wonder drive each one, and reports where it got the wrong answer even though every call succeeded. Pick Muse Wonder as the driver.
Muse Wonder sits above Muse Spark in Meta’s closed Muse line. It takes text, image, video, audio and PDF input, returns text, and has a 1M-token window. It launched alongside Muse Wonder Full, a stronger version that Meta keeps for Muse Code subscribers and does not serve through the API. Artificial Analysis scores Wonder 52 on its Intelligence Index, against 56 for Wonder Full and 58 for Claude Opus 5.5. Meta pitches two agent features: Active Memory, which carries context across conversations, and Active Focus, which keeps an agent working on its goal after it hits an obstacle.
These are the vendor’s own claims, not measurements of ours. Run the model against your own server to find out whether they hold for your tools.
Run the same four-step task on Muse Wonder and Muse Spark 1.2 against the complex-schema mock: create_user_profile, process_order, analyze_data, then configure_workflow. Spark 1.1 led MCP-Atlas, and Wonder has no MCP-Atlas score yet, so this is how you find out whether the bigger model calls your tools any better. All the mock servers →
| Model | Model ID |
|---|---|
| Muse Wonder | meta/muse-wonder |