Meta wins users over with ‘Muse’... bolsters tech edge with high-performance AI models [IT Item of the Day]

[Financial News] Meta Platforms has made a full-fledged push to compete with OpenAI and Anthropic, raising the performance of its in-house artificial intelligence (AI) models to the global top tier. Its personal AI agent, Muse, launched in September, has won users’ favor, and the company has now secured high-performance AI models to power it. Meta Platforms is expected to strengthen its AI development capabilities using its own super-sized data centers and step up its push into the AI agent market, where agents perform complex tasks on users’ behalf.
On the 12th (local time), Meta Platforms unveiled its next-generation AI models, Muse Wonder and Muse Wonder Full. The two models were developed by Meta Superintelligence Labs (MSL) using computing resources at Prometheus, its in-house gigawatt-scale data center. They are specialized in solving complex problems in fields such as law, finance and medicine, developing software, and carrying out tasks over extended periods.
The new products are seen as the next step in a strategy to take the performance of Muse, its personal AI agent, to the next level. Muse is a service that accesses various applications, including users’ email, calendars and shopping apps, and handles tasks such as sending messages and booking trips on their behalf. Since its launch in the United States, it has attracted users’ interest, with downloads exceeding 5 million over the month of September.
Meta Platforms had been operating the service based on its existing Muse Spark model, but has strengthened its competitive position by adding a high-performance model capable of handling more complex tasks. Meta Platforms CEO Mark Zuckerberg said, “If Muse Spark, which we introduced earlier, was our first model family, the much larger and more advanced Muse Wonder has taken us to the cutting edge.”
The models have also entered the global top tier in external performance evaluations. On the AI Intelligence Index (AAII) by AI performance analysis firm Artificial Analysis, Muse Wonder scored 52 points, while Muse Wonder Full, the higher-tier model, scored 56. Muse Wonder Full narrowed the gap to two points with Anthropic’s Claude Opus 5.5, which took first place with 58 points. The results are seen as evidence that Meta Platforms has established a competitive position in the high-performance AI model market, long led by OpenAI, Google and Anthropic.
Meta Platforms has particularly enhanced the models’ ability to remember users’ circumstances and see complex tasks through to completion. Active Memory retains users’ preferences and key information gleaned from previous conversations and uses them in later tasks. Active Focus enables the AI to continue working by finding another solution or correcting errors when unexpected problems arise during a task. For example, if a user asks for help filing an insurance claim or reviewing a lease, the AI analyzes relevant information and helps complete the necessary steps in sequence. The focus is on expanding the service beyond a chatbot that simply answers questions into an AI agent that performs real tasks.
Meta Platforms has begun making Muse Wonder available through a developer application programming interface (API) and its coding tool, Muse Code. The high-performance Muse Wonder Full model is available to Muse Code subscribers. Meta Platforms plans to add Muse Wonder to its personal AI agent Muse in the future, with paid subscribers getting access first.


[email protected] Choi Hye-rim Reporter