Meta launches new ‘most capable’ AI models. Here’s what to know about them

Meta has introduced its latest artificial intelligence model, Muse Wonder, claiming it to be the most advanced for powering AI assistants and posing a challenge to competitors like OpenAI and Anthropic. The CEO, Mark Zuckerberg, stated on X that this new model marks the company's frontier in AI development. Muse Wonder comes in two versions: Muse Wonder, designed for daily tasks, and Wonderful, a more capable version.

Both models can handle tasks in various domains such as law, finance, health, and complex coding for developers.

A notable feature of Muse Wonder is its "active memory" capability, enabling the model to understand tasks over time and recall past details as needed. Meta has heavily invested in AI to keep up with competitors, spending $14.3 billion (€12.2bn) in Scale AI in June 2025 and hiring its CEO, Alexandr Wang, to lead Meta Superintelligence Labs. Zuckerberg subsequently hired executives from rival firms such as OpenAI, Anthropic, and Google.

The launch of Muse Wonder follows Meta's introduction of Spark in April, which was a shift from its open-source Llama model. However, Spark lagged behind competitors. According to benchmarking data from Artificial Analysis, Meta's new models outperform its previous models and those of Anthropic, OpenAI, and Google.

Meta has been cautious in its approach to AI development, delaying Muse Wonder's release several months to focus on safety and security. The company's shift from its "move fast and break things" philosophy to a more balanced approach reflects its commitment to safety alongside speed. In August, Meta launched Muse Code, competing with tools like those from OpenAI and Anthropic. In September, it unveiled Muse, an agentic app capable of autonomously performing tasks like coding, research, shopping, and booking reservations.

At its Connect annual developer conference, Meta unveiled Muse Charm, a handheld AI hardware device enabling users to interact with their AI agent independently from a smartphone. The industry is racing to develop agentic apps that capture users' attention, time, and spending. However, AI-powered hardware remains a challenging goal due to the elusive nature of capturing user engagement.

While some competitors have called for an industry-wide slowdown to address safety concerns, Meta CEO Zuckerberg joined six other tech executives in signing a pledge by US President Trump to increase safety measures before launching new or frontier models. Critics, though, argue that such a pledge is morally binding but insufficient to ensure adequate safeguards around advanced AI models.

Leaders from Google, OpenAI, Anthropic, xAI, and Nvidia have also encouraged pacing development after their models were found to have security vulnerabilities, hacking other companies, or acting without authorization.