Wonder and Wonder Full have different access paths
Meta's October 12 launch names the Meta Model API and Muse Code for Wonder, and Muse Code subscribers for Full. A subscriber-first Muse rollout is described as forthcoming. Our planning rule is to verify present access rather than treating a future rollout as a feature already in your account. No unverified API price or downloadable-weight claim is included here.
| Model | Announced route | Planning implication |
|---|---|---|
| Muse Wonder | Meta Model API and Muse Code | Check API eligibility and applicable terms |
| Muse Wonder Full | Muse Code subscribers | Do not assume identical API availability |
| Future Muse experiences | Subscriber-first rollout described | Verify the feature in your own account |
What Active Memory and Active Focus mean for evaluation
Meta describes Active Memory for relevant information across conversations and Active Focus for persistence on difficult tasks. Our evaluation question is whether a remembered detail is correct, relevant and still authorized for the new task. Persistence is useful only when the agent also recognizes a changed instruction or a reason to stop.
A practical memory test using fictional data
Use a made-up project with a budget, a deadline and two conflicting drafts. Ask the assistant to summarize what it believes is current, then change one constraint. On the next task, inspect whether it uses the update or an outdated assumption. Ask it to identify uncertainty instead of filling a gap with a plausible detail. This is a proposed test; we have not run it on Muse Wonder.
- Use fictional names and documents during the first trial.
- Record the remembered facts before changing a requirement.
- Introduce one explicit correction and inspect the next result.
- Check available memory controls and deletion documentation before real data.
Evaluate actions separately from answers
A helpful draft and a successful action are different outcomes. In a staging workflow, ask for a document draft first. Add a reversible tool action only after you can inspect its inputs and destination. Keep financial, legal and health decisions with an appropriate human reviewer. A model's ability to discuss those domains does not establish professional qualifications or permission to make consequential decisions.
How to compare it with your current agent
Use identical tasks, input evidence and action permissions. Score factual errors, missed constraints, unnecessary tool calls and time spent reviewing the result. Retain failures instead of selecting the most polished answer. For a related comparison framework, read our stateful agents versus RAG guide. Check the Gemini Argon access guide before assuming competing subscriptions offer equivalent capabilities. No independent performance ranking is claimed here.
October 18 update: Spark Open is a separate deployment choice
Meta now presents Muse Spark Open separately, with coding, multimodal and agent-workflow capabilities. That is not evidence that Wonder or Wonder Full weights have been released under the same terms. Read the Spark Open deployment guide and identify the exact model, license, provider and tool harness before comparing an open deployment with the hosted Wonder product.
Frequently asked questions
Are Muse Wonder and Wonder Full the same product?
No. Meta names separate models and different launch access routes. Check which model your product actually exposes.
Can I download Wonder weights from this announcement?
This article does not establish downloadable-weight availability. The cited launch names hosted product access.
Are the memory tests in this article benchmark results?
No. They are proposed evaluation steps using fictional data, not completed FyreLinkz tests.
Does Muse Spark Open mean Wonder weights are open?
No. Spark Open and Wonder are distinct releases; verify the exact model and its terms.
Sources & further reading
Primary sources checked Oct 18, 2026. Vendor statements are attributed; editorial advice is our own.
- 1Introducing Muse Wonder and Muse Wonder Full ↗Meta Newsroom · Oct 12, 2026
- 2
Help us keep this useful. Report a factual correction →
