Meta claims new Muse Wonder models are world’s best at handling complex agentic tasks

Meta has introduced two new frontier models designed to carry out complex personal and professional tasks using long-term memory, sustained reasoning, and software tools.

The company says the models, Muse Wonder and Muse Wonder Full, deliver state-of-the-art performance across several evaluations covering business automation, coding, contextual memory, and professional work. The announcement did not provide numerical benchmark scores in the supplied text, making it difficult to independently assess the size of the performance gains.

Muse Wonder is available through the Meta Model API and Muse Code, while Muse Wonder Full is available to Muse Code subscribers. Meta plans to integrate Muse Wonder into its consumer assistant, initially for subscribers, and use it to generate digital artifacts and handle tasks requiring more advanced reasoning.

The release follows Meta’s introduction of Muse Spark and comes 14 months after the company began rebuilding its artificial intelligence infrastructure. The new models were trained at its gigawatt-scale Prometheus data center, reflecting Meta’s investment in larger computing systems and more capable models.

Memory guides complex tasks

A central feature of Muse Wonder is Active Memory, which allows the model to retain relevant information across conversations and retrieve it when needed.

Rather than requiring users to repeat earlier instructions, the system is designed to recognize when previously shared details matter to a new task. This could help personal assistants handle work involving preferences, documents and ongoing projects.

More from AI and Robotics

Meta also introduced Active Focus, a capability intended to keep the models working through difficult assignments. The system can try different approaches, respond to unexpected problems, and seek clarification when necessary. It is also designed to request confirmation before taking important actions.

The company evaluated these capabilities using MemoryBench for contextual recall and FocusBench for sustained task execution. It also used AutomationBench-AA to assess business workflows and professional evaluations covering legal and financial tasks.

Coding tests sustained reasoning

Software development is another major focus of the Muse Wonder series. Meta says both models can handle extended coding assignments involving multiple files, changing requirements, and repeated testing.

The company used SWE-Marathon and DeepSWE to evaluate performance on long-runningsoftwareengineering tasks. These assessments examine whether a model can maintain progress across complicated projects rather than simply produce isolated pieces of code.

Meta also tested how well its models perform on unfamiliar tasks. According to the company, Muse Wonder Full demonstrated stronger generalization than expected from the performance trend established by Muse Spark 1.3.

This suggests improvements may extend beyond the specific tasks emphasized during training, although the supplied announcement does not include enough numerical results to establish the size of that advantage.

Meta demonstrated Muse Wonder creating a playable 3D kitchen game from image inputs and building a personalized running plan through browser and application interactions.

The company also highlighted safeguards for handling private information, resisting malicious instructions, and limiting dangerous uses involving biological andcybersecurity threats.

Metaplans to use Muse Wonder to develop specialized models for its wider product ecosystem, although it has not announced a specific date for broader consumer availability.

Get the latest in engineering, tech, space & science - delivered daily to your inbox.

By subscribing, you agree to our

and

You may unsubscribe at any time.

0

COMMENT

Subscribe to

Today!

Access to exclusive content, expert insights and a deeper dive into engineering and tech. No ads, no limits.

By

With over a decade-long career in journalism, Neetika Walter has worked with The Economic Times, ANI, and Hindustan Times, covering politics, business, technology, and the clean energy sector. Passionate about contemporary culture, books, poetry, and storytelling, she brings depth and insight to her writing. When she isn’t chasing stories, she’s likely lost in a book or enjoying the company of her dogs.

• 1Scientists develop new catalyst for zinc-air batteries to run for over 300 hours
• 2US scientists grow better perovskite crystals, boost solar cell efficiency to 26%
• 3Video: Worm-like robot shifts its own weight to crawl over rocks, climb steps and swim
• 4Energy-importing countries record 3x faster clean energy generation than exporters
• 5Chinese EV giant's 31-joint humanoid robot could soon greet you at the car dealership

Meta has introduced two new frontier models designed to carry out complex personal and professional tasks using long-term memory, sustained reasoning, and software tools.

The company says the models, Muse Wonder and Muse Wonder Full, deliver state-of-the-art performance across several evaluations covering business automation, coding, contextual memory, and professional work. The announcement did not provide numerical benchmark scores in the supplied text, making it difficult to independently assess the size of the performance gains.

Muse Wonder is available through the Meta Model API and Muse Code, while Muse Wonder Full is available to Muse Code subscribers. Meta plans to integrate Muse Wonder into its consumer assistant, initially for subscribers, and use it to generate digital artifacts and handle tasks requiring more advanced reasoning.

The release follows Meta’s introduction of Muse Spark and comes 14 months after the company began rebuilding its artificial intelligence infrastructure. The new models were trained at its gigawatt-scale Prometheus data center, reflecting Meta’s investment in larger computing systems and more capable models.

Memory guides complex tasks

A central feature of Muse Wonder is Active Memory, which allows the model to retain relevant information across conversations and retrieve it when needed.

Rather than requiring users to repeat earlier instructions, the system is designed to recognize when previously shared details matter to a new task. This could help personal assistants handle work involving preferences, documents and ongoing projects.

More from AI and Robotics

Meta also introduced Active Focus, a capability intended to keep the models working through difficult assignments. The system can try different approaches, respond to unexpected problems, and seek clarification when necessary. It is also designed to request confirmation before taking important actions.

The company evaluated these capabilities using MemoryBench for contextual recall and FocusBench for sustained task execution. It also used AutomationBench-AA to assess business workflows and professional evaluations covering legal and financial tasks.

Coding tests sustained reasoning

Software development is another major focus of the Muse Wonder series. Meta says both models can handle extended coding assignments involving multiple files, changing requirements, and repeated testing.

The company used SWE-Marathon and DeepSWE to evaluate performance on long-runningsoftwareengineering tasks. These assessments examine whether a model can maintain progress across complicated projects rather than simply produce isolated pieces of code.

Meta also tested how well its models perform on unfamiliar tasks. According to the company, Muse Wonder Full demonstrated stronger generalization than expected from the performance trend established by Muse Spark 1.3.

This suggests improvements may extend beyond the specific tasks emphasized during training, although the supplied announcement does not include enough numerical results to establish the size of that advantage.

Meta demonstrated Muse Wonder creating a playable 3D kitchen game from image inputs and building a personalized running plan through browser and application interactions.

The company also highlighted safeguards for handling private information, resisting malicious instructions, and limiting dangerous uses involving biological andcybersecurity threats.

Metaplans to use Muse Wonder to develop specialized models for its wider product ecosystem, although it has not announced a specific date for broader consumer availability.

Get the latest in engineering, tech, space & science - delivered daily to your inbox.

By subscribing, you agree to our

and

You may unsubscribe at any time.

0

COMMENT

Subscribe to

Today!

Access to exclusive content, expert insights and a deeper dive into engineering and tech. No ads, no limits.

By

With over a decade-long career in journalism, Neetika Walter has worked with The Economic Times, ANI, and Hindustan Times, covering politics, business, technology, and the clean energy sector. Passionate about contemporary culture, books, poetry, and storytelling, she brings depth and insight to her writing. When she isn’t chasing stories, she’s likely lost in a book or enjoying the company of her dogs.

• 1Scientists develop new catalyst for zinc-air batteries to run for over 300 hours
• 2US scientists grow better perovskite crystals, boost solar cell efficiency to 26%
• 3Video: Worm-like robot shifts its own weight to crawl over rocks, climb steps and swim
• 4Energy-importing countries record 3x faster clean energy generation than exporters
• 5Chinese EV giant's 31-joint humanoid robot could soon greet you at the car dealership