Meta Launches Muse Wonder and Muse Wonder Full; Rebalancer Tutorial Details Solver Trade-offs

Both models offer a 1-million-token context window, accept text, image, video and audio as input, and output text. They run only on Meta's cloud through the API or the Muse Code CLI for macOS, Linux and Windows. Muse Wonder costs $4 per million input tokens, $20 per million output tokens and $0.20 per million cached tokens. Muse Wonder Full is subscription only and requires a Muse Code plan: Power at $50 per month or Maximum at $200 per month.

In Meta's published benchmark table at maximum reasoning effort, Muse Wonder Full leads 6 of 14 rows. It trails Claude Opus 5.5 on several measures. Muse Wonder Full scores 59.1 on Personal Superintelligence Index, 13.1 on Harvey LAB-AA, 59.1 on Vals Finance Agent V2, 45.4 on AutomationBench-AA, 78.4 on OSWorld V2.1 partial, 85.5 on MemoryBench, 57.1 on Terminal-Bench 4.0-AA, 79.0 on DeepSWE 1.1, 57.5 on SWE-Marathon at xhigh effort, 59.0 on ProgramBench and 56 on AA Intelligence Index v4. Muse Wonder scores 54.3 on Personal Superintelligence Index, 10.0 on Harvey LAB-AA, 56.6 on Vals Finance Agent V2, 39.1 on AutomationBench-AA, 72.6 on OSWorld V2.1 partial, 85.6 on MemoryBench, 52.0 on Terminal-Bench 4.0-AA, 78.8 on DeepSWE 1.1, 53.1 on SWE-Marathon, 50.0 on ProgramBench and 52 on AA Intelligence Index v4. Its best result is 85.6 on MemoryBench, the top score among the five models listed; its worst is 50.0 on ProgramBench, against 70.5 for Claude Opus 5.5. Meta notes that PSI, MemoryBench and FocusBench are internal evaluations, that competitor runs are best-effort and may not be tuned, and that SWE-Marathon results for rivals used Codex and Claude Code rather than a common harness.

Meta credits a larger base model for better results on unfamiliar tasks and says Muse Wonder was trained at its Prometheus data center with gains across pretraining and reinforcement learning. Two named behaviors define the release: Active Memory carries user context across conversations and retrieves details a task needs, while Active Focus keeps working on hard tasks, tries alternatives and checks in before important actions. For developers, the model plans long tasks and delegates to sub-agents, and it can operate browsers, desktop and mobile apps.

To use Muse Wonder, users open dev.meta.ai, sign in, add billing and create an API key, then export it as MODEL_API_KEY. The API is OpenAI-compatible and calls the Responses endpoint at https://api.meta.ai/v1/responses with the model name muse-wonder-1.0. Muse Code can be installed on Mac with curl -fsSL https://dev.meta.ai/install.sh | bash, on Windows with irm https://dev.meta.ai/install.ps1 | iex, and run with muse in a project folder. The default model is muse-spark-1.2, and users can switch with /models or launch with muse --model muse-wonder-1.0.

Separately, MarkTechPost published a tutorial on implementing Meta's Rebalancer assignment solver with a C++ core and Python binding. The tutorial describes objects, containers, dimensions, scopes and partitions, adds constraints and goals, and chooses between local search and the exact mixed-integer solver HiGHS. It builds a small datacenter, drains a busy host, and compares the two solvers on toy models, mutually exclusive hosts, synthetic clusters and 60 published OR-Library instances. The tutorial judges every plan by recounting the returned assignment rather than by the solver's flags. In this build, across 24 solves, the solved flag read True in 14 solves that still left memory over capacity. HARD under local search returned a violating start without raising; misspelled spec keys were ignored silently, and enum names were rejected. Timings came from an Apple M4 with the macOS wheel, and Colab's Linux wheel may time differently.