Meta introduces new AI models

Meta has introduced two new AI models, Muse Wonder and its more powerful version, Muse Wonder Full, following the mixed reception of Llama 4 in the spring of 2025. These models are designed to perform complex tasks independently, such as writing code or creating training plans, rather than just answering questions.

Models will be available to paying users

Muse Wonder will soon be available to the general public through the Muse application, but initially only to paying subscribers. Developers can already use it through the Meta Model API and in Muse Code, Meta's programming assistant. Muse Wonder Full is reserved for Muse Code subscribers.

Meta plans to integrate models into its personal agent

Meta plans to integrate these models into Muse, the personal agent it aims to incorporate into its glasses. The company also intends to use them to generate "artifacts" and develop specialized models for Muse users. To differentiate itself, Meta is emphasizing two key functions: Active Memory, which retains important user information across conversations, and Active Focus, which enables the agent to persist when stuck by trying other approaches, asking for clarification, and checking with the user before taking important actions.

Models trained in Meta's data centers

The models were trained in Prometheus, one of Meta's latest data centers. Meta claims first place in programming, law, finance, and health. For code, the group highlights SWE-Marathon, a test that measures a model's performance on long software projects. Muse Wonder Full achieved 57.5%, compared to 46.3% for Claude Opus 5.5 from Anthropic and 42.5% for GPT-6 Astra from OpenAI.

Independent rankings show close competition

On the Artificial Analysis index, an independent company that aggregates ten tests ranging from code to office work, Muse Wonder Full scores 56 points, about two points less than Claude Opus 5.5, which leads with 58, but ahead of GPT-6 Astra from OpenAI (53). The standard version of Muse Wonder caps at 52, and the previous Meta model, Muse Spark 1.3, at 48. The competition has not slowed down, with OpenAI launching GPT-6 Sol shortly after the release of Claude Opus 5.5, and Mistral playing the return card with Large 4.

Meta remains committed to open source

Meta has not stated whether it will publish the weights of Wonder, the parameters that allow anyone to run the model on their own devices. Alexandr Wang assures that the company remains "committed to open source," recalling that Meta recently published the weights of a smaller model from the Muse family. The full security report is pending, as Meta acknowledges that a personal agent acts on behalf of the user and can access their messages, calendar, or private documents.

Some users report unauthorized access

Some Muse users have complained on social media that the agent had read their personal messages without permission. According to Alexandr Wang, some users had given their consent earlier in the conversation but had forgotten it. For now, Wonder mainly concerns developers willing to pay for API access or a Muse Code subscription. Meta promises a complete security report "soon," without a date, and has not given any schedule for the model's arrival among free Muse users.