Meta ships Muse Wonder and Muse Wonder Full agentic models

Meta announced the Muse Wonder series on its research blog: two models, Muse Wonder and Muse Wonder Full. Meta calls them its most capable models, with state-of-the-art performance on personal and professional agentic tasks, and describes them as the next step on its scaling ladder and a major leap toward personal superintelligence. The pitch is that Muse Wonder brings together Active Memory and Active Focus, so the model keeps track of what matters to the user and works diligently toward what the user wants to accomplish.

Availability is split. Muse Wonder is available today through the Meta Model API and Muse Code. Muse Wonder Full is available today for Muse Code subscribers. Muse Wonder will soon power experiences in Muse and will roll out first to subscribers. It will also power artifact generation and processes that require more intelligence, and Meta says it will later use Muse Wonder to develop specialized models for everyone using Muse.

Meta frames the release as the result of a ground-up overhaul of its AI efforts that began fourteen months ago, moving from rebuilding its stack from first principles to pushing the frontier. It says the release builds on the approach of predictable, rigorous scaling it outlined with Muse Spark, across pretraining and reinforcement learning. Both models were trained at Meta's gigawatt-scale Prometheus data center.

The post claims state-of-the-art results across key benchmarks for personal and professional agents, memory and coding, with a results section comparing both models against other frontier models and a link to a report on the evaluations. It then looks closer at four areas. Domain expertise covers tasks in law, finance and health, such as understanding a lease or making sense of financial or health information; Harvey's Legal Agent Benchmark (LAB), which has models work from source documents and detailed instructions to produce legal work products, is one example. Active Memory means carrying important context about the user across conversations and actively retrieving the details relevant to the task. Active Focus means continuing toward a hard goal by trying alternative approaches, working through blockers and adapting to new information, while asking for clarification when needed and checking in before important actions; Meta evaluates it with FocusBench, and sustained execution with SWE-Marathon, which tests work on long-running software projects. Frontier coding ties these together: both models were trained on diverse, long-running tasks to sustain progress through requirements, changes across a codebase, and repeated cycles of testing and revision, from small applications to complex software.

A further section addresses why benchmark scores can mislead. Meta notes that two models can score alike while one handles a much wider range of unfamiliar tasks, which is what complaints of 'benchmaxxing' and praise for 'big model smell' both get at. To measure this, it plots Muse Spark 1.3 against Muse Wonder Full, each point being a distinct reasoning effort, with in-distribution score (tasks targeted in post-training) on one axis and average out-of-distribution score on the other. Both models improve similarly on out-of-distribution tasks as test-time compute scales, but given its in-distribution performance, Muse Wonder Full does better on unfamiliar tasks than the Muse Spark 1.3 trend would predict. Meta attributes the upward shift to larger, stronger base models, arguing that a stronger base makes post-training gains carry over more effectively.

On safety, Meta says a personal agent may handle messages, calendars and private documents, so it needs discretion with private information and resistance to misuse. It says Muse Wonder offers its strongest safeguards to date against CBRN (chemical, biological, radiological, and nuclear) and cyber misuse, while preserving access to frontier cyber capabilities for defensive use, per its Meta Superintelligence Scaling Framework. It calls Muse Wonder its most aligned model to date and says a full report with additional safety evaluations will be available soon.

Key facts

Meta introduced the Muse Wonder series: Muse Wonder and Muse Wonder Full, which it calls its most capable models for personal and professional agentic tasks.
Muse Wonder is available today via the Meta Model API and Muse Code; Muse Wonder Full is available today for Muse Code subscribers; Muse Wonder will soon power experiences in Muse, subscribers first.
Both models were trained at Meta's gigawatt-scale Prometheus data center, and Meta claims state-of-the-art results on key benchmarks for agents, memory and coding.
Meta says Muse Wonder Full does better on unfamiliar tasks than the Muse Spark 1.3 trend would predict, and credits larger, stronger base models.
Meta calls Muse Wonder its most aligned model to date, with its strongest CBRN and cyber safeguards so far; a full safety report is promised 'soon'.

Why it matters

This is Meta's newest step up the Muse line after Muse Spark, and the company positions it as frontier-level for agents, not just chat. The post's main bets are memory that persists across conversations (Active Memory) and sustained effort on hard, long-running work (Active Focus), measured with FocusBench and SWE-Marathon. Meta also argues that scores on popular benchmarks undersell breadth, and offers an in-distribution versus out-of-distribution comparison against Muse Spark 1.3 as its evidence that the stronger base model generalizes better.

Who it affects

Developers can reach Muse Wonder through the Meta Model API and Muse Code, and Muse Code subscribers also get Muse Wonder Full. People who use Muse will see Muse Wonder power experiences soon, with subscribers first in line. Meta's own examples of target work span law, finance and health tasks, and software projects from small applications to complex ones.

How to use it

Muse Wonder is available today through the Meta Model API and Muse Code. Muse Wonder Full is available today for Muse Code subscribers. Experiences in Muse come later: Meta says only 'soon', subscribers first. No pricing, API rates, subscription cost, context window or parameter counts are given in the text, and it does not explain how Muse Wonder differs from Muse Wonder Full.

How solid is it

This is a vendor announcement on Meta's own research blog, and the results are Meta's own evaluations; the text names no independent verification. The headline benchmark claims are not backed by numbers in the text: it contains no benchmark scores or numeric results, since the comparison tables and plots are not included in the text, and no competitor models are named. The generalization argument rests on one comparison against Muse Spark 1.3, Meta's own earlier model.

Risks and caveats

Meta itself notes that a personal agent may handle messages, calendars and private documents, which is why discretion and misuse resistance matter. Its safety claims ('strongest safeguards to date', 'most aligned model to date') are self-assessed, and the full report with additional safety evaluations is only promised 'soon', with no date. The Muse rollout is likewise undated. Muse Wonder Full is limited to Muse Code subscribers for now.

“Muse Wonder is our most aligned model to date, and alignment is central to how we build personal superintelligence.”

— Meta, Muse Wonder announcement