Technical Specifications
| Specification | Verified information |
|---|---|
| Model name | Muse Wonder Full |
| Developer | Meta |
| Announcement date | October 12, 2026 |
| Model family | Muse Wonder |
| Context window | 1,000,000 tokens |
| Primary capabilities | Coding, long-running agentic workflows, tool-assisted task execution |
| Streaming, function calling, and structured output parameters | Verify against the applicable product documentation |
Sources: Meta Muse Wonder model page, Meta launch announcement.
Important distinction: Muse Wonder has a public Meta Model API listing under muse-wonder-1.0, with published token pricing. Those details should not be copied to Muse Wonder Full; the two models have different access conditions.
What is Muse Wonder Full
Muse Wonder Full is a frontier AI model developed by Meta for complex coding and professional workflows that require planning, sustained reasoning, tool use, and error recovery. Announced on October 12, 2026, it is part of the Muse Wonder model family, alongside Muse Wonder.
The model is designed for work that extends beyond a single prompt or code completion. Its target tasks include making coordinated changes across large code repositories, handling extended software-engineering assignments, and supporting agentic workflows that execute multiple steps toward a defined objective. Meta also highlights Active Memory and Active Focus across the Muse Wonder family, describing capabilities intended to preserve relevant context and continue working when an approach encounters difficulties.
Benchmark Performance
Meta’s published model comparison reports the following results for Muse Wonder Full:
| Benchmark | Muse Wonder Full | What it evaluates |
|---|---|---|
| Personal Superintelligence Index | 59.1 | Personal and professional knowledge work |
| Harvey LAB-AA | 13.1 | Legal-agent tasks |
| Vals Finance Agent V2 | 59.1 | Financial research and analysis |
| HealthBench Pro | 62.3 | Open-ended health queries |
| AutomationBench-AA | 45.4 | End-to-end business workflows |
| GDP.pdf | 33.2 | Professional work |
| JobBench | 75.3 | Professional tool use |
| OSWorld V2.1 | 78.4 (partial) | Agentic computer use |
| MemoryBench | 39.4 (binary) | Context and memory retrieval |
| Terminal-Bench 4.0-AA | 57.1 | Agentic terminal coding |
| DeepSWE 1.1 | 79.0 | Long-horizon coding |
| SWE-Marathon | 57.5 (xhigh) | Extended software engineering |
| ProgramBench | 59.0 | Programming from scratch |
| Artificial Analysis Index | 56 | Composite model evaluation |
Source: Meta’s Muse Wonder model page.
What the results mean
Long-horizon software engineering is a key strength. The reported 79.0 on DeepSWE 1.1 and 57.5 on SWE-Marathon suggest that Muse Wonder Full is worth evaluating for repository-level changes, multi-stage engineering tasks, and workflows that require sustained execution.
Terminal-based coding is competitive, but not the top score in Meta’s comparison. Muse Wonder Full scored 57.1 on Terminal-Bench 4.0-AA, compared with 59.6 for Claude Opus 5.5 and 59.1 for GPT-6 Astra in the same published table. Model choice should depend on the specific workload rather than the model name or one benchmark alone.
Some benchmark labels require caution. OSWorld V2.1 is reported as a partial score, MemoryBench as a binary score, and SWE-Marathon at the xhigh reasoning setting. These labels should be retained when reproducing the figures. Scores across different benchmarks are not directly comparable.
These are published evaluation results, not guarantees of performance on a specific production codebase. Before selecting a model, test it against representative tasks, including patch correctness, regression rate, tool-call reliability, completion time, and total workflow cost.
Use Cases
Muse Wonder Full is most relevant for teams evaluating advanced agentic coding and multi-step task execution:
- Large repository engineering: Coordinated changes across multiple files, debugging, and refactoring.
- Extended software development: Tasks that require iterative implementation, testing, and recovery from failed approaches.
- Agent-based development: Workflows in which a model plans subtasks and operates through an execution environment.
- Complex programming assignments: Building functionality from a specification rather than completing a small code fragment.
- Benchmark-driven model selection: Evaluating long-horizon coding performance alongside terminal-based execution and general task performance.
These are suitable evaluation scenarios based on the model’s published positioning, not a guarantee that every workflow is supported by every interface.