Meta Unveils First Frontier Model, 'Muse Wonder,' Joining Top-Tier AI Race
IT DAILY · · 1 views
✦ AI Summary
Meta unveiled its first frontier AI models, Muse Wonder and Muse Wonder Full, on the 12th.
The two models were developed based on the existing Muse Spark, with pretraining and reinforcement learning applied more extensively.
Muse Wonder Full scored 45.4 points on AutomationBench-AA, 59.1 points on Vals Finance Agent V2, and 56 points on Artificial Analysis's Intelligence Index.
Meta unveiled its first frontier AI models on the 12th local time. The models released this time are Muse Wonder and Muse Wonder Full. Both models were developed based on the existing Muse Spark, with pretraining and reinforcement learning applied more extensively.
For training infrastructure, Meta used the gigawatt-scale data center Prometheus. Meta said the models delivered top-tier performance on major benchmarks for personal and work agents, memory, and coding.
According to disclosed figures, Muse Wonder Full scored 45.4 points on AutomationBench-AA, which evaluates automation capabilities across work tasks. That topped GPT-6 Astra's 41.6 points and Claude Opus 5.5's 41.2 points.
Muse Wonder Full also scored 59.1 points on Vals Finance Agent V2, which evaluates financial analysis and research capabilities. The comparison models cited were Claude Fable 5.1 at 58.9 points and GPT-6 Astra at 53.5 points. Attention is now on whether Meta can prove its competitiveness in the global AI model race with these models.
Meta unveiled Muse Wonder and Muse Wonder Full. On Artificial Analysis's Intelligence Index benchmark, Muse Wonder Full scored 56 points. The model tied for second place with Claude Sonnet 5.5, while the top spot went to Claude Opus 5.5 with 58 points.
Muse Wonder scored 52 points. Muse Wonder tied for seventh place with GPT-6 Sol.
The two models support interpretation of legal, financial, and health-related information, as well as a range of tasks including complex software development. Meta said the models are based on larger and more powerful foundation models and have also improved their ability to respond to tasks not heavily covered during training.
Meta emphasized the ability to remember context that matters to users and to keep working toward goals. To that end, Meta said it implemented Active Memory and Active Focus. Active Memory maintains context across conversations and retrieves information needed for tasks, while Active Focus is a function that keeps work going by trying other methods when obstacles arise, Meta said.
Muse Wonder is priced at USD 4 per 1 million input tokens and USD 20 per 1 million output tokens. Meta described this as an important technological advance toward bringing personal superintelligence to everyone. It also said it looks forward to helping more people realize their dreams and create the world they want.
Along with this case from Meta, major AI companies have also recently been releasing frontier models in succession.
Anthropic released Claude Opus 5.5 on the 22nd of last month, and OpenAI released GPT-6.1 Sol on the 29th. Google announced Gemini 4 Argon on the 30th and also began offering Gemini 4 Argon first to verified cybersecurity partners.
Source: IT DAILY · Yang Seung-gap
Original: https://www.itdaily.kr/news/articleView.html?idxno=242160
References
This article was produced with the help of an automated content generation algorithm.
Source: IT DAILY
View originalThis article was summarized and organized by BizCrush based on the original article from IT DAILY. For exact quotations and full details, please refer to the original article.