Meta Unveils Muse Wonder Series, Its Most Capable AI Models Yet

Meta on October 12, 2026 announced the Muse Wonder series, introducing Muse Wonder and Muse Wonder Full as frontier models for personal and professional agentic tasks. Meta described them as its most capable models to date, built to deliver state-of-the-art performance on work ranging from navigating an insurance claim to understanding a lease to building a working app.

Capabilities for Agentic Tasks

According to the announcement, the Muse Wonder models bring capability to expert domains including law, finance, and health, along with frontier coding performance for people building everything from a game to an app for a new business. Meta framed the release as arriving fourteen months after it began an overhaul of its AI efforts, a period in which it went from rebuilding its foundations to pushing the frontier. The models build on groundwork laid with Muse Spark.

Meta built the new models around features it calls Active Memory and Active Focus. Active Memory carries context about a user across conversations and brings back the details relevant to the task at hand. Active Focus keeps the model working toward a goal when a task gets hard: trying other approaches, recovering from errors, asking for clarification when needed, and checking in before following through on anything important.

Coding ties those capabilities together. Meta’s research blog describes building software as a process of tracking requirements, making changes across a codebase, and working through repeated cycles of testing and revision, and says both models were trained on diverse, long-running tasks to sustain progress across that process. Demonstrations in the post show Muse Wonder building a playable 3D kitchen game from customizable image inputs, assembling a running plan in a browser using OpenStreetMap data based on personal preferences and circumstances, designing and ordering custom swag from real photos taken from a personal Instagram profile, and preparing a custom wedding seating chart from a list of complex matching requirements.

Training and Generalization

Muse Wonder and Muse Wonder Full were trained at Meta’s gigawatt-scale Prometheus data center. The release builds on the scaling approach Meta outlined with Muse Spark, applied across pretraining and reinforcement learning.

The research blog also examines how performance on tasks targeted during post-training relates to performance on unfamiliar tasks as test-time compute scales, a breadth of capability it says popular benchmarks in a few domains cannot fully capture. Meta’s analysis compares Muse Spark 1.3 and Muse Wonder Full, plotting each model’s score on a composite representative of tasks targeted during post-training (in-distribution) against average performance on tasks not targeted during post-training (out-of-distribution). Meta reported that both models improve similarly on out-of-distribution tasks as test-time compute scales, but that, given its in-distribution performance, Muse Wonder Full performs better on unfamiliar tasks than the Muse Spark 1.3 trend would predict. Meta attributes that upward shift to the larger, stronger base models behind the Muse Wonder series, whose post-training gains carry over more effectively to unfamiliar tasks, versatility the blog says matters for personal agents that must handle many combinations of goals, tools, and constraints.

Evaluation Suite

Meta said the Muse Wonder series achieves state-of-the-art results across key benchmarks for personal and professional agents, memory, and coding, and published a full results table comparing the models with other frontier models across a broad evaluation suite. The named evaluations include Harvey’s Legal Agent Benchmark, which requires models to work with source documents and follow detailed instructions to produce legal work products, and Vals Finance Agent V2 for financial-agent tasks. Long-term contextual memory is measured on MemoryBench, focus through FocusBench, sustained execution on long-running software projects through SWE-Marathon, and long-horizon coding through DeepSWE. Meta pointed to a separate report for additional evaluation details.

Safety and Alignment

Meta said it trained the Muse Wonder family to avoid sharing private information without permission, to turn down requests to assist in performing actually harmful acts, and to work to improve security and strengthen defenses against cyberattacks. Because developers and security teams can use models like Muse Wonder to find flaws in their own code and strengthen their systems, Meta said it made sure the models can still do that work.

Meta described Muse Wonder as offering its strongest safeguards to date against CBRN (chemical, biological, radiological, and nuclear) and cyber misuse while preserving access to frontier cyber capabilities for defensive use, and called Muse Wonder its most aligned model to date. The safety evaluations described in the research blog cover unauthorized code execution attack success rates, personal-information handling violation rates, biological-risk refusal rates, and cyber misuse refusal and cyber defense compliance rates. Meta trains, evaluates, and releases its most capable models in line with its Meta Superintelligence Scaling Framework, and said a full report with additional safety evaluations will be available soon.

Availability and Stated Plans

Muse Wonder is available through the Meta Model API and Muse Code, while Muse Wonder Full is available for Muse Code subscribers. Meta said Muse Wonder will soon power experiences in Muse, rolling out first to subscribers, and that it will also power artifact generation and other processes that require more intelligence. Meta also said it will use Muse Wonder to develop specialized models it can bring to everyone using Muse.