TL;DR
Get audio and creator gear delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
Meta has introduced Muse Wonder and Muse Wonder Full, two models it says are designed for personal and professional agent tasks, coding and memory. Muse Wonder is available through the Meta Model API and Muse Code; Wonder Full is available to Muse Code subscribers. Meta has not yet published independent validation of its performance claims in the supplied material.
Meta has announced Muse Wonder and Muse Wonder Full, two AI models it says are built for personal and professional agent tasks, including coding and work that requires retaining context over time. Meta says Muse Wonder is available through its Meta Model API and Muse Code, while Wonder Full is currently available to Muse Code subscribers; integration into the Muse product is planned.
Meta describes the models as its most capable to date and says they combine two features it calls Active Memory and Active Focus. Active Memory is intended to carry relevant information across conversations and retrieve it for later tasks. Active Focus refers to sustained work toward a goal, including trying different approaches, adapting to new information and asking for clarification or checking with a user before important actions, according to Meta.
The company says the models were trained at its Prometheus data center and developed through work on pretraining and reinforcement learning. Meta reports strong results across evaluations involving legal, financial and health-related tasks, coding, memory and agent performance. It points to benchmarks including Harvey’s Legal Agent Benchmark, FocusBench and SWE-Marathon, which tests progress on long-running software projects.
Meta also argues that performance can carry over to tasks outside those specifically targeted during post-training. In its comparison of Muse Wonder Full with Muse Spark 1.3, the company says Wonder Full scored better on unfamiliar tasks than the earlier model’s trend would predict when compared at similar in-distribution performance. These are company-reported evaluations; the supplied announcement does not provide independent confirmation of the results.
A Push Toward Persistent AI Agents
The announcement matters because Meta is positioning the models not simply as chat tools, but as agents that can retain useful context and work through multi-step tasks. That approach could support tasks such as building software, organizing plans or handling work that spans multiple conversations and tools. Whether it makes those experiences more dependable in daily use will depend on how well the models follow instructions, handle errors and seek permission before consequential actions.
Availability is also staged. Developers can access Muse Wonder through the Meta Model API, and users of Muse Code can use Wonder; the more capable Wonder Full version is limited to Muse Code subscribers, according to Meta. The company says the models will later power experiences in Muse, initially for subscribers, as well as artifact generation and processes that need greater intelligence. This gives the release a product role beyond a research announcement, while leaving the timing and exact scope of broader access open.
Meta’s emphasis on generalization is relevant to agent products because users’ goals, available tools and constraints vary widely. A model that performs well only on a narrow set of prepared tests may struggle when a task differs from training examples. Meta says its results suggest Wonder Full handles some unfamiliar tasks better than expected from its performance on targeted evaluations, but readers should treat that as Meta’s interpretation of its testing, not a settled independent finding.
As an affiliate, we earn on qualifying purchases.
Meta’s Model-Scaling Strategy
Meta says the Wonder release builds on a broader reorganization of its AI work that began 14 months before the announcement. The company describes that effort as a ground-up overhaul, moving from rebuilding its technical stack to scaling training and pushing model capabilities. It previously introduced Muse Spark and presented an approach to predictable scaling; Meta says the Wonder models extend that work.
The announcement frames Wonder as part of Meta’s stated goal of bringing what it calls personal superintelligence to users of its products. That phrase is the company’s ambition, not a demonstrated outcome. The stated roadmap is to bring Wonder into Muse, first for subscribers, use it for artifact generation and other demanding processes, and develop specialized models for wider use later. The supplied material does not give dates for those later stages.
Meta describes its evidence through a suite of benchmarks and demonstrations, including a playable 3D kitchen game built from customizable image inputs and an agent demonstration involving browser and app use. Such examples illustrate intended capabilities, but do not establish how consistently the models perform across real-world users, tasks or safety-sensitive situations.
“Muse Wonder brings together Active Memory and Active Focus to keep track of what matters to you and work diligently toward what you want to accomplish.”
— Meta
As an affiliate, we earn on qualifying purchases.
Independent Testing And Rollout
The announcement’s performance claims are based on Meta’s own evaluations. The supplied material does not include independent test results, enough detail to verify every comparison, or a full account of how the models perform across everyday use. It is also unclear how the models handle errors in high-stakes legal, financial or health-related tasks; benchmark performance alone does not establish suitability for professional decisions.
Meta has not specified when Muse Wonder will reach all Muse users, how subscription access will change, or which features will be available in each product. The description of Active Memory also leaves practical questions unanswered, including how users control what is retained, how long information persists and what safeguards apply when personal context is retrieved. Details about permissions and oversight for actions taken through agents are likewise limited in the supplied report.
As an affiliate, we earn on qualifying purchases.
Access Expands In Stages
For now, Meta says users can access Muse Wonder through the Meta Model API and Muse Code, while Muse Code subscribers can use Muse Wonder Full. The company says Wonder will later power Muse experiences, with subscribers receiving access first, and will also support artifact generation and other processes requiring more intelligence.
Meta has not announced dates for those product rollouts in the supplied material. Further information to watch for includes access terms, demonstrations of the models’ behavior in real tasks, evaluation methods and user controls for memory and agent actions. Meta also says it plans to use Wonder to develop specialized models for broader Muse users, but has not provided a timeline or described what those models will do.
As an affiliate, we earn on qualifying purchases.
Key Questions
What are Muse Wonder and Muse Wonder Full?
They are two AI models announced by Meta. The company says they target personal and professional agent tasks, including coding, sustained work and retrieving relevant information from earlier conversations.
Where can people use the models now?
Meta says Muse Wonder is available through the Meta Model API and Muse Code. Muse Wonder Full is available to Muse Code subscribers.
When will Muse Wonder be available in Muse?
Meta says Wonder will soon power experiences in Muse and that rollout will begin with subscribers. The announcement does not give a specific date.
Are the performance claims independently verified?
The announcement reports results from Meta’s evaluations and names several benchmarks. The supplied material does not provide independent validation of those claims.
What does Active Memory mean?
Meta uses the term for carrying important context across conversations and retrieving details it considers relevant to a later task. The announcement does not fully explain user controls or retention settings.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
