What Muse Wonder Benchmarks Don't Tell You: The Full Model Isn't Open to Developers

Meta has returned to the frontier-model contest with Muse Wonder and the more capable Muse Wonder Full. But the provocative question in the headline needs an immediate qualification: the standard model is available through Meta’s model API, while the Full version is initially limited to Muse Code subscribers. Neither is presented as a downloadable open-weight release.

That does not prove Meta has abandoned open models. It does show that the company’s most ambitious new agent product follows a more controlled, commercial distribution strategy than the Llama releases that built Meta’s open-model reputation.

What Meta actually released

Muse Wonder accepts text, images, video, and audio. Meta emphasizes two agent-oriented capabilities: Active Memory, which can retrieve relevant information across interactions, and Active Focus, which helps the system change tactics and continue a difficult task rather than stopping at the first failure.

The company positions it for long coding sessions and high-stakes professional work, including legal, financial, and health-related tasks. Those claims raise the bar for evidence. At launch, much of the performance case comes from Meta’s own evaluations, and the detailed safety evidence available to outsiders remains more limited than the breadth of the use cases being promoted.

The distribution split matters. Developers can access regular Wonder through Meta’s API, but Wonder Full—the model associated with the strongest claims—is tied to the paid Muse Code experience. There is no general downloadable weight package or broadly available Full API at launch.

Why the missing download matters

Open weights allow researchers and companies to inspect, fine-tune, host, and govern a model on their own infrastructure. A hosted API offers convenience, but the provider controls access, pricing, logging, feature changes, and shutdown risk.

For an agent that remembers documents, messages, calendars, and long-running tasks, deployment control is not a philosophical detail. It determines where sensitive context is stored, who can audit access, and whether an organization can delete or migrate its operational memory.

Meta may eventually expand access. Until it does, buyers should evaluate Muse Wonder as a hosted commercial service, not assume that the Llama playbook automatically applies.

Is Meta’s open-model strategy over?

The evidence supports a narrower conclusion: Meta is running two strategies at once. Open-weight releases can build an ecosystem and push standards outward. Closed frontier services can capture subscription revenue, protect expensive capabilities, and give Meta tighter control over safety and abuse.

That hybrid approach is common across the industry. The important change is symbolic. Meta’s strongest newly announced capability is being sold through a controlled product before it is offered as infrastructure other developers can freely deploy.

What developers should compare

Before switching from Claude, OpenAI, or another provider, teams need more than a benchmark chart:

1. Access: Is the required model available by API, subscription product, or downloadable weights?
2. Price: What are input, output, memory, and tool-use costs for long agent sessions?
3. Memory control: Can users inspect, correct, export, and delete stored information?
4. Approval boundaries: Which external actions require confirmation?
5. Auditability: Can an organization reconstruct what the agent saw and did?
6. Independent evidence: Do outside tests reproduce the provider’s performance claims?

Verdict

Muse Wonder is important because it signals Meta’s return to high-end agent competition. The sharper story, however, is distribution. Meta’s strongest model is not an open download and is not generally available through an unrestricted API.

The open-model era at Meta is not necessarily ending. But openness is no longer the default assumption for its most valuable frontier capability.