Meta unveils Muse Wonder family and launches a powerful model for knowledge tasks

Meta has launched a new family of artificial intelligence models called Muse Wonder, which it described as the largest and most capable in its history, placing it alongside companies such as OpenAI, Anthropic and Google in the race to develop advanced models.

The family consists of two main versions: Muse Wonder, geared toward everyday tasks, and the larger, higher-capacity Wonder Full, designed specifically for knowledge applications that rely on smart agents, scientific reasoning and programming.

CEO Mark Zuckerberg said the new model "represents a larger, more advanced class of models that moves us to the forefront," and Meta said it relied on independent evaluation data showing the new models outperform its previous generation known as Spark and approach the performance of competing companies' models.

Meta highlighted two new features in these models: "active focus" and "active memory," and said these capabilities help smart agents maintain conversational context and follow multi-step professional tasks through to completion.

The company said the models are immediately available to developers via its API, and that it will soon support paid services within the Muse app.

The announcement comes weeks after Meta revealed a new portable device that enables interaction with the smart agent independently of the phone, amid an intense race among tech giants to develop agents capable of performing tasks such as search, shopping and bookings on users' behalf.