Meta releases Muse Wonder models for personal and professional agents
Meta has introduced Muse Wonder and Muse Wonder Full, but has not disclosed how developers can access, license or price the agent models.
By Wei-Lin Zhao · AI Correspondent
· 3 min read
Meta has released the Muse Wonder models, a two-model family it says is designed for personal and professional AI agents. The company announced Muse Wonder and Muse Wonder Full on Oct. 12, but did not disclose pricing, licensing, API access, model sizes or whether either model is already running its consumer-facing Muse agent.
On its AI site, Meta calls both models “frontier” systems and claims state-of-the-art results on agent tasks. That is Meta’s own performance characterization, rather than an independent assessment.
What are Meta Muse Wonder models designed to do?
Meta is aiming the models at extended, multistep work delegated to an agent, according to SiliconANGLE’s account of the launch. The company cited tasks such as handling an insurance claim and named coding, law, finance and health among areas where it says the models are capable.
The release adds two named features. Meta says Active Memory can retain relevant details from prior conversations and surface them later, while Active Focus is intended to help an agent continue when it hits an error or roadblock. Under Meta’s description, Active Focus may try alternative approaches and seek clarification or user permission for consequential actions.
SiliconANGLE reported that the models can accept mixed text, video, audio and image inputs in a context window of up to 1 million tokens, while producing text-only responses. The supplied material does not specify the models’ hardware requirements or safety-evaluation results.
Muse Wonder is distinct from Meta’s Muse agent
The new model family should not be confused with Muse, Meta’s personal AI agent that the company announced in September. Meta’s own site presents Muse as a separate consumer product and Muse Wonder as its latest model release. Meta previously said Muse was powered by its earlier Muse Spark model; it has not confirmed in the material available whether Muse Wonder has replaced it.
Meta says the consumer agent can work across connected applications from a dedicated cloud virtual machine, use a browser, fill out forms and continue some long-running work after a user closes the app. It says Muse asks for approval before actions such as sending email or making purchases. Those are stated capabilities of Muse, not a confirmed list of public Muse Wonder functions.
The distinction is consequential for teams evaluating the announcement. Meta has announced underlying models and their intended agent use, but has not established where businesses or developers can obtain them, on what terms, or what production controls they provide.
Meta compared Muse Wonder Full with models from OpenAI and Anthropic in its model card, SiliconANGLE reported. The comparison includes Meta’s warning that its runs of competing systems were best-effort and may not have used optimal prompts or tools, so the figures are not independent validation.
This story draws on original reporting from SiliconANGLE.
