Muse Wonder

A frontier example that leads on individual and expert agents

Today, we’re enthusiastic to current the Muse Wonder sequence alongside Muse Wonder and Muse Wonder Full, our most capable models that provision state-of-the-art achievement on individual and expert agentic tasks. Together, they grade the next stage on our scaling ladder and a important jump toward individual superintelligence. Muse Wonder brings together Active Memory and Active Focus to keep track of what matters to you and activity diligently toward what you desire to accomplish.

Muse Wonder is accessible today through the Meta Model API and Muse Code, and Muse Wonder Full is accessible today for Muse Code subscribers.

Muse Wonder volition shortly power experiences in Muse, and volition rotate out archetypal to subscribers. Muse Wonder volition additionally power relic generation and processes that necessitate additional intelligence. And in the future, we’ll be using Muse Wonder to create specialized models that we can bring to everyone using Muse.

Scaling Toward Personal Superintelligence

Fourteen months ago, we began a ground-up overhaul of our synthetic intellect efforts. In that time, we’ve gone from rebuilding our stack from archetypal principles to pushing the frontier.

When we introduced Muse Spark, we outlined our method to predictable, rigorous scaling. Today’s publish builds on that activity throughout pretraining and reinforcement learning, delivering our most capable models yet. Trained at our gigawatt-scale Prometheus data center, Muse Wonder and Muse Wonder Full bring together the activity of teams throughout Meta to provision frontier capabilities for complex tasks at activity and in mundane existence — another stage toward bringing individual superintelligence to the billions of group who use our products.

Frontier Capabilities

Model Highlights

The Muse Wonder sequence achieves state-of-the-art results throughout key benchmarks for individual and expert agents, memory, and coding.

The complete results below difference Muse Wonder and Muse Wonder Full alongside another frontier models throughout a broad evaluation suite.

For additional particulars concerning our evaluations, see our report.

The results complete provision a broad perspective of the models’ performance. We now obtain a nearer appearance at four areas chief to their activity on real-world tasks: domain expertise, Active Memory, Active Focus, and frontier coding.

Domain Expertise

Muse Wonder and Muse Wonder Full use specialized cognition to problems group encounter at activity and in regular life, from understanding a lease to making awareness of financial or health information. Our evaluations contain tasks in law, finance, and health that test this ability. Harvey’s Legal Agent Benchmark (LAB), for example, requires models to activity alongside origin documents and prosecute detailed instructions to create lawful activity products.

Active Memory

One of the most crucial steps towards helpful individual and expert agents is dependable memory. Models need to recall what matters and acknowledge whenever before data becomes relevant. We call this Active Memory: carrying crucial environment concerning you throughout conversations and actively retrieving the particulars applicable to the project at hand.

Active Focus

Using that environment to complete a project additionally requires sustained effort. Active Focus describes how Muse Wonder continues operating toward a challenging goal: trying substitute approaches, operating through blockers, and adapting to new information. Along the way, it asks for clarification whenever needed and checks in before taking crucial actions. We measure concentration through FocusBench and sustained implementation through SWE-Marathon, which tests activity on long-running application projects.

Frontier Coding

Coding brings these capabilities together. Building application requires tracking requirements, making changes throughout a codebase, and operating through repeated cycles of evaluation and revision. Muse Wonder and Muse Wonder Full provision frontier coding performance, supporting projects from small applications to complex software. We trained the two models on diverse, long-running tasks to assistance them prolong advancement throughout this process.

Capabilities That Generalize

Two models frequently have akin benchmark scores, yet one seems to grip a much wider range of unfamiliar tasks. That’s what complaints of “benchmaxxing” and praise for “big example smell” the two get at: a breadth of capability that famous benchmarks in a few domains can’t completely capture.

To quantify this mismatch, we examine how achievement on tasks targeted during post-training relates to achievement on unfamiliar tasks as we measure test-time compute. The storyline below compares Muse Spark 1.3 and Muse Wonder Full, alongside all item denoting achievement at a distinct reasoning effort. The horizontal axis shows the model’s mark on a composite delegate of the tasks we focused on during post-training (“in-distribution”). The vertical axis shows the average mark on a set of tasks we did not mark during post-training (“out-of-distribution”).

Both of our models enhance likewise on out-of-distribution tasks as we measure test-time compute. But stated its in-distribution performance, Muse Wonder Full performs equal improved on unfamiliar tasks than the Muse Spark 1.3 trend would predict. This upward change is since Muse Wonder and Muse Wonder Full build on larger, stronger basis models: a stronger basis yields post-training gains that transport complete additional efficiently to unfamiliar tasks. That versatility particularly matters for individual agents, which need to grip many combinations of goals, tools, and constraints — not all of which we can anticipate during training.

Safety at the Frontier

Delivering individual superintelligence to everyone requires a important funding in alignment and safety. A individual delegate takes actions on a user’s behalf and power grip environment specified as messages, calendars, and personal documents. That flat of admission demands a example that exercises discretion alongside personal data and resists misuse. We trained the Muse Wonder family to guide on the two dimensions, alongside robustness to adversarial instructions and careful handling of personal information.

As models rotate into additional capable, our goal is to assistance group harness those capabilities safely and responsibly. Muse Wonder offers our strongest safeguards to date against CBRN (chemical, biological, radiological, and nuclear) and cyber misuse, during preserving admission to frontier cyber capabilities for protective use, per our Meta Superintelligence Scaling Framework.

Muse Wonder is our most aligned example to date, and alignment is chief to how we build individual superintelligence. A complete study alongside additional safety evaluations volition be accessible soon.

We anticipation that Muse Wonder and Muse Wonder Full provision additional group the power to prosecute their dreams and build the earth they want. This is a important specialized stage toward individual superintelligence for everyone, alongside much additional to come.