Meta launches Muse Wonder and Wonder Full for Muse Code subscribers

Meta has released two models at once, Muse Wonder and Muse Wonder Full, and claims they can go toe to toe with the best in the industry. Since October 12, both have been available to Muse Code subscribers, and Wonder can also be accessed through the Meta Model API. This is already the third notable release in the Muse lineup in a short time. There are no figures confirmed by independent testing yet, so the bold claim remains just a claim.

What exactly was released

The models continue a chain of releases that Meta has been building since relaunching its Superintelligence Labs division. The company first showed Muse Spark, the first model from the revamped lab. Then came the Muse AI agent, which generated a lot of buzz. Now Wonder and Wonder Full have been added.

Access is uneven. Muse Code subscribers get both models, so Meta is handing them first to developers and people who write code with AI. Only Wonder is available via API. The announcement says nothing about that option for Wonder Full, and that is an important detail: the senior model currently lives inside the product and has not become a building block for third-party applications.

The available materials do not say how Wonder and Wonder Full differ in size, context window or inference cost. Meta calls them competitors to the industry's best models, but offers no specific tests or comparisons.

Where this is heading

The mechanics are the most interesting part for now. According to Meta, the Muse agent will eventually run on Wonder, starting with subscribers. After that, the company wants to build specialized models that it "can offer to everyone who uses Muse." Meta has not disclosed a list of such models or a schedule.

The rollout order is typical for the market. First comes the paying audience, which is more tolerant of rough edges and generates a steady stream of feedback. Mass-market users come later. For Meta, which The Verge describes as catching up in the AI race, this is a way to show progress without the risk that a model's weak spot becomes known to hundreds of millions of people.

The bet on coding is also easy to read. Agentic programming is currently what brings AI companies the clearest revenue and the fastest feedback. The fact that Muse Code, rather than a free chatbot, became the first platform for Wonder says something about priorities.

What is known and what is not

• Two models were released: Muse Wonder and Muse Wonder Full.
• Muse Code: both models are available, and subscribers received them on the day of the announcement.
• Meta Model API: only Wonder is available.
• Plans: power the Muse agent with Wonder, first for subscribers, then build specialized models for all Muse users.
• Not disclosed: model parameters, context window size, pricing, benchmark results, the timeline for moving the agent to Wonder.

What to test in practice

If you already pay for Muse Code, it makes sense to run Wonder and Wonder Full on your own work tasks: refactoring, tests, getting to grips with an unfamiliar repository. Compare the results with what you use now, on the same set of prompts. That is the only way to check a claim of parity with the leaders.

A team considering the Meta Model API should not build production on it yet. Without public pricing, limits and latency data, you cannot work out the economics, and the absence of Wonder Full from the API makes it impossible to judge the quality ceiling. A sensible step now is to prototype on Wonder and wait for independent tests. If Meta does move the Muse agent to Wonder, that is where it will become clear how well the model holds up through multistep work, rather than single responses.