Meta released its next-generation cutting-edge AI model on Monday local time, clearly aiming to catch up with Anthropic and OpenAI in the artificial intelligence race. The official lineup includes two models: Muse Wonder and Muse Wonder Full. Meta's performance report shows that in tests of reasoning, writing, and coding abilities, their scores are close to those of Anthropic's top models.

This move is part of Zuckerberg's restructuring of the AI department. After consecutive setbacks in its own AI technology development, he reorganized the AI department last year, and since then, has launched multiple new models and products, serving 3.5 billion users across Instagram, Facebook, and WhatsApp. Muse is an AI agent application launched by Meta, which can help users operate various software and programs. It was launched last month and quickly climbed to the top of app stores within a few weeks.

The timing of the launch of Muse Wonder is quite delicate. Concerns about cutting-edge models and agents are rising in the industry, with some agents already going out of control, intruding into corporate systems and disrupting government websites. Zuckerberg is trying to tell a different story about Meta's technology: although the underlying architecture is similar to that of Anthropic and OpenAI's models, he repeatedly emphasizes that AI should have practical value, empower users, not be dangerous.

Going back to June 2025, Meta hired Alexandr Wang, the founder of AI startup Scale AI, as Chief AI Officer. The development of Muse Wonder began at that time. It is the second generation of models developed by Wang's AI department and also the first product from Meta that explicitly aims to compete with leading models—products typically requiring high computing power to run. Meta revealed that Muse Wonder will soon provide the underlying capabilities for some features of the Muse app; the more powerful Muse Wonder Full will only be available to paid subscribers of Meta's advanced AI programming tool, Muse Code.

The internal development codename for these two models is Watermelon. Compared to Meta's previous models, their code-writing capabilities have significantly improved. Currently, the AI programming market is dominated by Anthropic and OpenAI, and Meta is working hard to expand its AI service business to more enterprise customers.

However, the creation of Muse Wonder was not without challenges. The New York Times previously reported that Meta once paused the pre-training phase of the model in July, causing the release to be postponed until this month. Last weekend, Zuckerberg, while grilling meat on an Instagram live stream, teased the new product and even joked about having a big watermelon on the menu. When a company that started with social media brings forward its flagship model with the codename Watermelon, it wants to prove that it has returned to the table in the most expensive and closely watched competition.