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 describes as its most capable to date for personal and professional agent tasks. Muse Wonder is available through the Meta Model API and Muse Code; Muse Wonder Full is available to Muse Code subscribers. Meta has not yet provided independent verification of its performance claims or detailed terms for its planned rollout in Muse.
Meta has introduced Muse Wonder and Muse Wonder Full, a pair of AI models the company says are designed for personal and professional agent tasks, including coding and work that requires sustained effort. Meta says Muse Wonder is available through its Model API and Muse Code, while Wonder Full is available to Muse Code subscribers; the company plans to bring Wonder to its Muse product later.
Meta describes the models as its most capable so far and says they were developed through additional work in pretraining and reinforcement learning. The company says both models were trained at its Prometheus data center and are intended to handle complex tasks across work and everyday life. Those descriptions and performance claims come from Meta; the announcement does not provide independent evaluation of them.
The company highlights two features it calls Active Memory and Active Focus. Active Memory refers to retaining relevant information across conversations and retrieving it for later tasks, according to Meta. Active Focus describes the models’ ability to try different approaches, respond to new information and ask for clarification. Meta says the models check in before important actions, but the announcement does not specify how that safeguard works across different tools or uses.
Meta says the models perform well on evaluations covering law, finance, health and coding, as well as agent tasks involving memory and sustained work. It cites Harvey’s Legal Agent Benchmark, FocusBench and SWE-Marathon among the evaluations discussed. The company also describes a demonstration in which Muse Wonder creates a playable 3D kitchen game from customizable images. The announcement points to a broader evaluation report, but benchmark results depend on test design and do not establish how reliably the models will perform in everyday use.
Where Meta Plans to Use Wonder
The release puts Meta’s latest models into developer and coding workflows while setting up a possible wider consumer rollout. If the company’s claims hold up in practical use, features such as cross-conversation memory and sustained task execution could let an AI agent handle work that involves several steps, changing requirements or information from previous conversations. That could matter for users who want help with projects rather than one-off answers.
For now, availability is split: Wonder is offered through Meta’s API and Muse Code, while Wonder Full is limited to Muse Code subscribers. Meta says Wonder will soon power experiences in Muse, initially for subscribers, and could support artifact generation and processes requiring more intelligence. Those plans make the announcement relevant to both developers testing the models and users waiting to see whether the capabilities appear in Meta’s consumer-facing products.
The practical test will be reliability, not just benchmark performance. Agents that retain personal context or act across applications may be more useful, but also raise questions about what information is remembered, when it is retrieved and how consequential actions are controlled. Meta’s announcement describes the intended features but leaves those implementation details open.
As an affiliate, we earn on qualifying purchases.
Meta’s Scaling Program
Meta frames the release as part of a broader rebuilding of its AI effort. The company says it began a ground-up overhaul 14 months before the announcement and previously outlined a scaling approach when it introduced Muse Spark. It presents Muse Wonder as a continuation of that work, rather than as a standalone product update.
The announcement also argues that models need to handle unfamiliar combinations of goals, tools and constraints, not only tasks represented in standard tests. Meta compares Muse Wonder Full with Muse Spark 1.3 on in-distribution and out-of-distribution evaluations, saying Wonder Full performs better on unfamiliar tasks than its earlier model’s trend would predict. This is Meta’s interpretation of its own evaluation; the release does not establish how the comparison would hold across independent tests or all real-world tasks.
Meta says it plans to use Wonder to develop specialized models for Muse users. The company does not give a schedule for that work. Its announcement therefore covers both current access for developers and subscribers, and a longer-term plan to use the model family in consumer experiences.
“Muse Wonder and Muse Wonder Full are our most capable models.”
— Meta, in its Muse Wonder announcement
Open Questions on Access and Results
The announcement does not set out independent benchmark results, detailed model limitations or how the evaluations compare with competitors under identical conditions. Meta refers readers to a separate evaluation report, but benchmark outcomes alone cannot show whether the models will complete complex tasks consistently outside test settings.
It is also not clear when Muse Wonder will reach Muse, which subscribers will receive it first, or what features will be available at launch. Meta has not specified how long Active Memory retains information, what controls users will have over remembered details, or how the models’ checks before important actions will operate. The scope of access, pricing and availability by region is not established in the announcement.
The Planned Muse Rollout
Meta says Muse Wonder will soon power experiences in Muse, with subscribers receiving access first. It also says the model will support artifact generation and processes that require more intelligence, and that it intends to develop specialized models for Muse users. The company has not announced specific dates for those steps.
Developers can currently access Muse Wonder through the Meta Model API and Muse Code, according to Meta; Muse Code subscribers can access Wonder Full. The next useful indicators will be the timing and scope of the Muse rollout, further technical information from Meta, and independent testing of the models’ performance on long-running tasks and unfamiliar requests.
Key Questions
What did Meta announce?
Meta announced Muse Wonder and Muse Wonder Full, AI models it says are designed for personal and professional agent tasks, including coding and sustained work.
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.
What are Active Memory and Active Focus?
Meta describes Active Memory as keeping and retrieving relevant context across conversations. Active Focus refers to continuing work through obstacles, adapting to new information and asking for clarification when needed.
When will Muse Wonder be available in Muse?
Meta says Muse Wonder will come to Muse soon and roll out first to subscribers, but it has not announced a specific date or detailed eligibility.
Are the performance claims independently verified?
The announcement presents Meta’s own benchmark results and descriptions. It does not provide independent verification of the claims or establish how the models will perform across all real-world tasks.
Source: hn
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
