Meta* releases Muse Wonder and Muse Wonder Full: two models for AI agents with memory

Meta has released two models at once: Muse Wonder and Muse Wonder Full. The company calls them the most powerful in its history. Both were designed for AI agents, meaning systems that work through a chain of actions rather than a single reply. The base Muse Wonder is already available through the Meta Model API and inside the Muse Code coding agent. The higher-tier Muse Wonder Full is open only to Muse Code subscribers for now. The headline of the Habr publication says the models outperformed GPT-6 Astra and Gemini 4 Argon. The source text itself contains no test results, so for now this is a claim, not a verified fact.

What exactly was released

Meta*'s key phrasing is that the models are built for "personal and professional" agents. That implies three properties. First, the model remembers the user between conversations and doesn't start from a blank slate every time. Second, it doesn't stop at the first error but tries a different path. Third, it carries a long task through to the end without losing the thread at step ten.

These are the places where agents most often break in practice. Modern models answer short questions well. The harder case is different: forty steps in a repository, a test that breaks on step thirty-two, and a context that is already half lost by then.

Two models, two doors

The access scheme shows how Meta* sets its priorities. Muse Wonder goes both to developers through the API and to Muse Code users. Muse Wonder Full stays inside the company's own product and is limited to paying subscribers. This lets the company road-test the strongest version on a limited audience before releasing it through an external API.

Muse Code, meanwhile, becomes a showcase. A coding agent offers the most visible test of the claimed resilience: either the task is solved and the tests pass, or it isn't.

What is known so far

• Models: Muse Wonder and Muse Wonder Full.
• Positioning: the most powerful models in Meta*'s history.
• Target scenario: personal and professional AI agents.
• Claimed properties: memory between conversations, continuing work after errors, completing long tasks.
• Muse Wonder: available through the Meta* Model API and in Muse Code.
• Muse Wonder Full: only for Muse Code subscribers.
• Comparison with GPT-6 Astra and Gemini 4 Argon: mentioned in the source's headline; the text contains no benchmark data.

What to check before adopting

If you build agents, you don't have to wait for headline numbers: you can evaluate the model on your own tasks right now through the Model API. Take five to ten real long-running scenarios where your current model falls apart, and run them without hints. Look at the share of tasks completed end to end, not the quality of individual answers. Check memory separately: what exactly the model remembers, where that data is stored, and whether it can be erased, especially if conversations may contain personal or corporate data.

It's worth clarifying pricing and limits right away. The original publication doesn't state them, and for agents that take dozens of steps per task, they determine the economics. For now, only Muse Code subscribers can evaluate Muse Wonder Full; everyone else will have to wait for independent tests.

*Meta has been designated an extremist organization and is banned in Russia.