Meta Unveils New AI Models 'Muse Wonder' to Chase OpenAI and Anthropic
TECHWORLD · · 3 views
✦ AI Summary
Meta unveiled new AI models, "Muse Wonder" and "Muse Wonder Full," aimed at strengthening AI agents' task execution capabilities.
The two models come with conversation-memory features and error-handling capabilities during tasks, and were developed by Meta Superintelligence Labs (MSL) using the computing infrastructure of Prometheus, Meta's in-house data center.
Meta will offer Muse Wonder through Meta Model API and Muse Code, while Muse Wonder Full will be provided to paid Muse Code subscribers.
Meta unveiled its new AI models, "Muse Wonder" and "Muse Wonder Full." Both models are designed to strengthen AI agents' task execution capabilities. On the 12th local time, Meta announced the launch of the two models and began offering them through developer APIs and coding tools.
The two models come with memory features for conversation content and mechanisms to handle errors during tasks. Meta introduced them as part of a move to catch up with leading AI companies such as OpenAI and Anthropic.
The models were developed by Meta Superintelligence Labs (MSL). Their development drew on computing infrastructure from Meta's in-house data center, Prometheus.
Compared with the existing Muse Spark, the two models are larger in scale and were trained on more data. Based on that, Meta released the two new models to improve AI agents' practical capabilities.
Alexandr Wang, Meta's CAIO, described the two models as state-of-the-art in an interview with Bloomberg News. He also expressed confidence that they can compete with market-leading models.
According to Artificial Analysis, Muse Wonder scored 52 on the AI intelligence index (AAII). Muse Wonder Full scored 56, tying for second place with Anthropic's Claude Sonnet 5.5. Claude Opus 5.5 ranked first with 58, leaving Muse Wonder Full 2 points behind Claude Opus 5.5. Meanwhile, OpenAI GPT-6 Astra and Google Gemini 4 Argon each scored 53, putting Muse Wonder Full 3 points ahead of GPT-6 Astra and Gemini 4 Argon.
Meta outlined plans to use the new models for AI agents. Target tasks include coding and specialized work in law, finance and medicine. Major use cases include checking insurance claims procedures, analyzing contracts and reviewing financial information, all of which involve multiple stages of judgment and processing.
Meta said it applied Active Memory and Active Focus for that purpose. Active Memory remembers key information from earlier conversations and user preferences, then uses that information and those preferences in later tasks. As a result, users do not need to repeat the same explanations and can continue working while reflecting the context of previous conversations.
Active Focus was designed to prompt different approaches when errors occur during complex tasks or when problems are difficult to solve using existing methods. The feature also asks users for additional information when needed and is designed to require confirmation before carrying out important tasks.
These two features are used in multi-step software development. The AI writes, revises and tests code while retaining previously provided requirements, and then searches for new ways to solve problems when issues arise.
Meta used the Harvey LAB and Vals Finance Agent V2 to verify the models' task execution capabilities. It also used MemoryBench to evaluate long-term memory and FocusBench to assess task persistence.
Meta has set different delivery methods for each model. Muse Wonder is available through Meta Model API and Muse Code. The higher-end Muse Wonder Full is provided to paid Muse Code subscribers.
Meta plans to apply the new models to Muse, its in-house AI agent application. The company aims to expand the use of the models from developer tools to consumer services, supporting document writing, information analysis and software development.
Because AI agents access a wide range of information, including personal documents and schedules, Meta also emphasized improved safety. The company said it trained the models to prevent the sharing of sensitive information without user consent and also improved protections against privacy leaks and cyberattacks.
Meta plans to expand its AI services business by offering the new models to developers and individual users. At the same time, the company intends to improve model performance and focus on applying them to real-world work and services.
Meta CEO Mark Zuckerberg, while describing Muse Spark and Muse Wonder in contrast, said that Muse Spark was assumed to be the first model lineup, while Muse Wonder was larger in scale and more advanced, and that it had pushed Meta to the cutting edge of technology.
Source: TECHWORLD · Kim Seung-gi
Original: https://www.epnc.co.kr/news/articleView.html?idxno=408136
References
This article was produced with the help of an automated content generation algorithm.
Source: TECHWORLD
View originalThis article was summarized and organized by BizCrush based on the original article from TECHWORLD. For exact quotations and full details, please refer to the original article.