Two models, three Muse products

At launch, the standard Muse Wonder model is available through the Meta Model API and Muse Code. Muse Wonder Full, the larger model, is limited to Muse Code subscribers. Muse Wonder will soon power experiences in Muse and will roll out first to subscribers.

Meta’s naming spans the Muse Wonder model family, the Muse Code development product, and the Muse personal agent. Developers evaluating the release should distinguish the underlying model from the products that provide access to it.

Prometheus supplies the compute

Meta says the release follows a 14-month overhaul of its AI stack that began with Muse Spark. The company trained the new models at its gigawatt-scale Prometheus data center using a combination of pretraining and reinforcement learning, a post-training process that rewards useful model behavior.

The company reports that Muse Wonder Full ranks first on its Personal Superintelligence composite, ahead of Claude Opus 5.5, GPT-6 Astra, and Claude Fable 5.1. Because Meta reports these results, the ranking requires external verification. Its practical meaning also depends on the composite’s task mix, weighting, prompts, tool access, and inference budget.

Memory and persistence drive the pitch

Meta uses the term Active Memory for retaining user context across conversations and retrieving relevant details for a new task. A personal agent could use that capability to remember preferences, previous decisions, or project history without receiving the same background in every prompt.

Active Focus describes continued work toward a goal through alternative approaches, blocker resolution, clarification requests, and approval checks before consequential actions. In a coding workflow, that could mean inspecting a failed test, changing the implementation, running the suite again, and asking for guidance when the requirements remain ambiguous.

Persistent memory introduces product and security requirements beyond recall accuracy. Production deployments need clear controls for consent, retention, deletion, tenant isolation, sensitive-data filtering, and the context exposed to each tool call.

Meta’s benchmark case, with caveats

Meta reports gains in memory, sustained software work, coding, professional knowledge, and misuse resistance. Lower scores are better only for the two violation and attack metrics identified below.

Meta also cites evaluations in law, finance, and health, including Harvey’s Legal Agent Benchmark and Vals Finance. These results indicate the kinds of professional tasks the company tested, but they do not establish accuracy for a specific jurisdiction, institution, patient population, or production workflow.

Cross-vendor comparisons remain sensitive to model versions, prompts, tool configurations, sampling settings, time limits, and inference budgets. Teams should reproduce representative tasks under equal conditions before using these figures for procurement or architecture decisions.

Generalization rises with inference compute

Meta’s generalization experiment separates tasks represented during post-training from unfamiliar tasks outside that distribution. It then increases test-time compute, meaning the inference effort available after a model receives a task, which can include additional reasoning steps or attempts.

In Meta’s reported results, both Muse Wonder models improve on unfamiliar tasks as test-time compute increases. Muse Wonder Full also exceeds the trend predicted from Muse Spark 1.3’s performance, which Meta attributes to a stronger pretrained base model whose post-training gains transfer more effectively.

An externally reproducible version of this experiment would help determine whether the gains extend across task families and inference settings. Useful disclosures would include the task-selection process, contamination checks, compute budgets, stopping rules, scoring method, and variance across runs.

Safety scores leave systems work

Because the planned Muse agent may interact with calendars, messages, documents, and developer tools, its permissions and failure modes carry direct consequences. Meta reports lower attack and privacy-violation rates than selected competitors, alongside stronger refusal behavior for chemical, biological, radiological, nuclear, and cyber misuse.

The company says its cyber safeguards preserve access to defensive capabilities under the scaling framework. The reported 94.5% cyber defense compliance score, compared with 15.7% for GPT-6 Astra, is large enough that independent testing and full methodological details are necessary for interpretation.

Model-level safeguards cover only part of an agent deployment. Applications still need least-privilege credentials, tool allowlists, sandboxing, confirmation gates, audit logs, prompt-injection defenses, transaction limits, and recovery procedures for incomplete or duplicated actions.

Before Muse touches real tools

• Reproduce long-task performance. Test repository-scale work with the same time, tool, and token budgets used for incumbent models.
• Measure recovery behavior. Introduce failed tool calls, stale state, ambiguous instructions, and partial outages to see whether the model recovers safely.
• Audit memory controls. Verify what the system stores, how it retrieves context, how users delete it, and whether information can cross accounts or tenants.
• Benchmark cost and latency. Longer inference can improve results while increasing response time and spend. Measure completed-task cost, not token price alone.
• Test authorization boundaries. Confirm that model requests cannot bypass application permissions, approval steps, or environment isolation.
• Check operational details. Confirm current pricing, context limits, rate limits, regional availability, data-retention terms, model versioning, and service-level commitments in Meta’s API documentation.

Muse Wonder Full’s reported SWE-Marathon lead gives coding-agent teams a concrete result to investigate, while Active Memory offers a relevant capability for assistants that span many conversations. Adoption will depend on whether those gains survive production constraints, especially equal-budget testing, privacy controls, tool security, latency, and cost.