The Pulse
Meta Introduces Muse Wonder Models for Agents and Coding
Meta introduced Muse Wonder and Muse Wonder Full on October 12, describing them as models for personal and professional agent tasks. The company says the models combine persistent memory with sustained task work, and lists API and Muse Code
AI.info Team · Written with AI ·
“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 AI Research, introducing the Muse Wonder models
Meta introduced Muse Wonder and Muse Wonder Full on October 12, presenting the pair as its most capable models for personal and professional AI-agent tasks. The release adds a task-focused branch to the Muse family, with features intended to help the models retain useful details between conversations and keep working toward a goal when they encounter obstacles.
Muse Wonder is available through Meta Model API and Muse Code. Muse Wonder Full is available to Muse Code subscribers. Meta says Muse Wonder will soon power experiences in its Muse assistant, rolling out first to subscribers, as well as artifact generation and processes that require more intelligence.
Memory and persistence are the pitch
Meta calls the two main features Active Memory and Active Focus. Active Memory carries important user context across conversations and retrieves details relevant to a later task. Active Focus describes the model’s effort to keep working toward a challenging goal by trying alternative approaches, working through blockers and adapting to new information. The company says the model can also ask for clarification and check in before important actions.
Meta points to tasks such as interpreting a lease and making sense of financial or health information. Its research post describes evaluations in law, finance and health, alongside coding tests that include long-running software projects. Examples shown in the post include using a browser and apps to build a running plan, designing custom merchandise from personal photos and preparing a wedding seating chart.
Benchmark comparisons use different setups
Meta’s model card lists text, image, video and audio as supported inputs, text as the output, and a one-million-token context length at inference. It says Muse Wonder and Muse Wonder Full results were generated with maximum reasoning effort, as were comparisons with the listed OpenAI and Anthropic models in the overall methodology. There is an exception: for SWE-Marathon, the card reports Muse Wonder Full at xhigh reasoning effort.
The card also describes differences in how some coding comparisons were assembled. External results retain the harness or agent product used by the reporting vendor or model provider. Internally generated competitor runs use common settings where practical, but Meta characterizes them as best-effort, noting that prompts, tools and runtime may not be optimally tuned for proprietary models. The performance claims are company-reported results across multiple evaluations, rather than a single uniform test.
Safety assessments and limitations
Meta says the models were trained to handle private information carefully and resist misuse, while retaining access to frontier cybersecurity capabilities for defensive use. The model card says Meta does not rule out Muse Wonder Full reaching its high-risk threshold in the chemical and biological domain without mitigations. The company says it applied safety post-training and system-level classifiers; with mitigations and safeguards in place, it assessed risk as moderate or lower across the domains it evaluated.
The card also lists limitations: the model may give inaccurate, biased or objectionable responses, and may make errors in multi-step reasoning, particularly in unfamiliar situations. Visual and spatial reasoning, especially on 3D tasks, remains a known limitation. Meta recommends that developers assess the model for their own uses and consider safeguards such as human confirmation before irreversible actions.
Launch access centers on Meta’s products
The announcement describes access through Meta’s developer API and Muse Code, with Muse Wonder Full limited to Muse Code subscribers. Meta also says it plans to bring Muse Wonder to its assistant, beginning with subscribers. The release and model card describe these access routes and model capabilities; they do not announce an open-weight release.
For now, Meta’s case for the new models rests on its benchmark results, product access and stated safety measures. The launch distinguishes between Muse Wonder, available through the API and Muse Code, and Muse Wonder Full, offered to Muse Code subscribers.