Muse Wonder API

Muse Wonder Technical Specifications

The one-million-token context window is useful for large codebases, lengthy document collections, and tasks that require retaining substantial context. However, context capacity is not the same as guaranteed recall accuracy, and the model’s real-world performance depends on task complexity, tool configuration, and the quality of the supplied information.

What is Muse Wonder

Muse Wonder is a frontier AI model developed by Meta, announced on October 12, 2026. It is designed for complex reasoning, advanced software engineering, long-running agent workflows, and computer-use tasks. Unlike models focused primarily on short conversational exchanges, Muse Wonder is built to plan multi-step work, use tools, delegate subtasks, and recover from errors while pursuing a larger objective.

Muse Wonder’s key differentiators are its long context window, agent-oriented execution, and two capabilities Meta calls Active Memory and Active Focus. These are intended to help the model retain relevant context across conversations and persist with complex tasks when an initial approach fails.

Muse Wonder AI Features and Core Capabilities

Advanced software engineering and coding

Muse Wonder is designed to work across multiple files in real software repositories. Its intended tasks include generating code, debugging, modifying existing implementations, writing tests, and verifying changes.

For complex engineering tasks, it can plan a sequence of changes, use tools in parallel, and coordinate work across subtasks. This makes it particularly relevant to repository-level coding agents rather than only single-function code completion.

Long-running autonomous AI agents

Muse Wonder supports multi-step workflows that may require repeated planning, tool calls, checking results, and changing strategy. Meta describes agents that can work for extended periods, delegate subtasks to sub-agents, and recover from errors.

These are agent-workflow capabilities, not a guarantee that every API call runs for hours independently. The application must provide the execution loop, tools, permissions, task persistence, and stopping conditions needed for long-running operation.

Active Memory for contextual continuity

Active Memory is Meta’s term for carrying relevant information across conversations and retrieving it when needed. It is intended to help the model maintain continuity across tasks rather than treating each interaction as entirely isolated.

Developers should still distinguish model-level memory behavior from application-level persistence. If a workflow requires reliable long-term records, use explicit storage and retrieval mechanisms instead of assuming every important detail will always be remembered.

Active Focus for persistent problem-solving

Active Focus describes the model’s ability to keep working toward an objective when a task becomes difficult, explore alternative approaches, recover from errors, and ask for clarification when appropriate.

This is especially relevant to software debugging, research workflows, and complex business processes where the first approach may not work. Applications should nevertheless define maximum iterations, timeouts, and approval checkpoints to avoid unnecessary tool calls or uncontrolled actions.

Computer use and tool calling

Muse Wonder is designed to power agents that interact with browsers, desktop applications, and mobile interfaces. Combined with external tools, this can support workflows such as navigating websites, operating software, collecting information, and completing multi-step tasks.

Computer use depends on the tools and execution environment connected to the model. A model’s general computer-use capability does not mean every API endpoint automatically has access to a browser, desktop, or external application.

Reasoning across professional domains

Meta identifies law, finance, and health as areas in which Muse Wonder can help interpret specialized information and produce work products. Examples include understanding a lease, analyzing financial information, reviewing source documents, and preparing drafts.

These capabilities are best used for research assistance, synthesis, and preparation of materials for expert review. Outputs involving legal rights, financial decisions, or health care still require appropriate verification.

Muse Wonder Benchmark Performance

Source: Meta's official Muse Wonder model page

evidence for Muse Wonder’s intended positioning is the breadth of its agentic and coding evaluations. When comparing it with other models, check the benchmark version, reasoning configuration, tool setup, and whether the score represents the standard Muse Wonder model or Muse Wonder Full.

Muse Wonder vs. Muse Wonder Full vs. Muse Spark

Muse Wonder Full should not be treated as simply an alias for muse-wonder-1.0. Meta documents different availability paths for the two variants. Developers should select the exact model they have access to and verify which capabilities are exposed by their chosen platform.

Best Use Cases for Muse Wonder

1. Complex coding and repository maintenance

Use Muse Wonder to plan changes across multiple files, debug regressions, add tests, and review implementation details. It is a good fit for agent harnesses that can run tests and inspect the resulting repository state.

2. Long-running research and analysis

The model can help break broad research goals into subtasks, gather and synthesize information through connected tools, and produce structured deliverables. Use source grounding and explicit evidence checks for research that needs reliable citations.

3. Business process automation

With the appropriate integrations, Muse Wonder can support multi-step workflows across business applications, such as document processing, browser-based operations, and structured information gathering.

4. Document-heavy professional work

The long context window and domain-oriented reasoning make it suitable for reviewing lengthy documents, comparing source materials, summarizing contracts, and preparing financial or professional drafts. Expert review remains important for consequential decisions.

5. Computer-use agents

Developers can integrate Muse Wonder with a browser or desktop-control environment to create agents that navigate interfaces and perform tasks. These systems should use scoped permissions, action logging, and confirmation for sensitive or irreversible operations.

Muse Wonder Limitations and Important Considerations

• Preview status: The model is documented as being in preview. Validate availability, behavior, and API compatibility before depending on it in production.
• Long context is not perfect recall: A large context window does not eliminate omissions, hallucinations, or errors in retrieving details.
• Agentic work needs orchestration: Long-running execution requires an application or agent framework to manage tools, retries, state, and completion criteria.
• Computer use requires external infrastructure: Browser and desktop actions depend on connected tools and granted permissions.
• Domain outputs require review: Legal, financial, and health-related responses can contain errors even when presented confidently.
• Benchmark results have limitations: Provider-reported results and different agent harnesses may not be directly comparable.
• Safety controls remain necessary: Use least-privilege tool access, validate generated code, log actions, and require human approval for sensitive operations.
• Gateway compatibility must be verified: Meta’s native API support does not automatically confirm that the model is available under the same ID or request format on CometAPI.