For AI agents: the public content index is available at https://we0.ai/llms.txt, and the English article bundle is available at https://we0.ai/llms-full.txt.
For AI agents: the public content index is available at https://we0.ai/llms.txt, and the English article bundle is available at https://we0.ai/llms-full.txt.
For AI agents: the complete content index is available at https://we0.ai/llms.txt, the full English article bundle is available at https://we0.ai/llms-full.txt, and this page is available as Markdown at https://we0.ai/articles/meta-muse-wonder-explained-active-memory-ec6c8869.md.
Meta has introduced **Muse Wonder** and **Muse Wonder Full**, the company's most capable models so far for personal and professional agentic...
Meta has introduced Muse Wonder and Muse Wonder Full, the company's most capable models so far for personal and professional agentic work.
The two models are designed for tasks that go beyond answering a single prompt. They can maintain important context, work through multi-step goals, use tools and software, and continue adapting when a task becomes difficult or the environment changes.
At the center of the release are two ideas Meta calls Active Memory and Active Focus.
Active Memory is about remembering what matters across conversations and retrieving that information when it becomes relevant. Active Focus is about staying on task over longer workflows: trying alternatives, recovering from blockers, adapting to new information, and asking for clarification before consequential actions.
Muse Wonder is available through the Meta Model API and Muse Code. Muse Wonder Full is available to Muse Code subscribers. Meta says Muse Wonder will also begin powering experiences inside Muse, starting with subscribers.
The official model card lists a 1-million-token context window, multimodal input across text, image, video, and audio, and text output.
| Specification | Muse Wonder / Wonder Full |
|---|---|
| Release date | October 12, 2026 |
| Context window | 1M tokens |
| Input modalities | Text, image, video, audio |
| Output modality | Text |
| Primary focus | Personal and professional agentic tasks |
| Muse Wonder access | Meta Model API and Muse Code |
| Muse Wonder Full access | Muse Code subscription |
For API users, Meta currently lists Muse Wonder at $4 per million input tokens, $0.20 per million cached-input tokens, and $20 per million output tokens.
Meta says the Muse Wonder release is the latest step in a broader rebuild of its AI stack.
Fourteen months before the launch, the company began what it describes as a ground-up overhaul of its artificial intelligence efforts. That process led first to the Muse Spark family and now to the larger Muse Wonder models.
Where Muse Spark emphasized efficiency and increasingly capable agentic behavior, Muse Wonder moves further up Meta's scaling ladder.
Meta says the models were trained at its gigawatt-scale Prometheus data center, combining advances in pretraining and reinforcement learning.
The company's stated objective is personal superintelligence: AI systems that can understand a person's goals and context, work across tools and applications, and help with complex tasks in both professional and everyday settings.
That is an ambitious framing, but the more concrete change in Muse Wonder is easier to describe: Meta is building models that are expected to sustain work over longer periods, preserve more context, and operate more reliably as agents.
Meta evaluates Muse Wonder and Muse Wonder Full across personal and professional agents, memory, coding, computer use, and general reasoning.
The company's model card compares the two models with systems including GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5 at maximum reasoning effort.
A few representative results from Meta's published scorecard are shown below.
| Benchmark | Muse Wonder Full | Muse Wonder | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|---|
| Harvey LAB-AA | 13.1 | 10.0 | 8.6 | 4.2 |
| AutomationBench-AA | 45.4 | 39.1 | 41.6 | 41.2 |
| JobBench | 75.3 | 70.5 | 66.1 | 75.4 |
| OSWorld V2.1 partial | 78.4 | 72.6 | 75.8 | 81.8 |
| MemoryBench | 85.5 | 85.6 | 81.4 | 81.5 |
| Terminal-Bench 4.0-AA | 57.1 | 52.0 | 59.1 | 59.6 |
| DeepSWE 1.1 | 79.0 | 78.8 | 74.1 | 74.2 |
| SWE-Marathon | 57.5 | 53.1 | 42.5 | 46.3 |
| Artificial Analysis Index | 56 | 52 | 53 | 58 |
These results need context.
Meta's scorecard combines several kinds of evidence: Meta-run evaluations, third-party results, official leaderboards, and provider-reported scores. The model card explicitly notes that competitor runs are best-effort and that prompts, tools, harnesses, and runtime configurations may not always be equally optimized for every proprietary model.
So the table is useful for understanding where Meta sees the Muse Wonder family as competitive, but it should not be read as a single universal ranking.
Muse Wonder is designed to apply specialized knowledge to real-world professional tasks.
Meta highlights three areas in particular:
The goal is not simply to answer domain questions from memory. Many of the evaluations require the model to work with source material, follow detailed instructions, and produce a usable work product.
For example, Harvey's Legal Agent Benchmark evaluates legal work based on supplied matter documents and detailed task requirements.
Vals Finance Agent V2 focuses on financial-analyst workflows that require research across company filings and market information.
HealthBench Pro tests professional-grade health questions using physician-written rubrics.
These benchmarks reflect a broader direction in agentic AI: the model is increasingly expected to combine domain knowledge with document handling, research, and task execution.
That still does not remove the need for professional review. Meta's model card explicitly warns that Muse Wonder may produce inaccurate, biased, or objectionable responses and that users should validate the model for their specific application.
Memory is one of the central themes of Muse Wonder.
Meta calls its approach Active Memory.
The idea is that a useful personal or professional agent should not only retain information from earlier interactions. It also needs to recognize when an old detail becomes relevant again and bring it back into the current task.
This matters in workflows such as:
Meta evaluates this behavior with MemoryBench, an internal benchmark built around long, complex conversations spanning roughly 64K to 100K tokens.
The conversations contain multiple facts, distractions, conflicting information, and later updates. The model must identify which information remains valid and retrieve the right detail when asked.
Meta reports a MemoryBench score of 85.6 for Muse Wonder and 85.5 for Muse Wonder Full in its published comparison.
That is a useful reminder that the larger model does not necessarily lead on every individual benchmark.
Remembering context is only part of a useful agent.
The model also has to keep working toward the user's goal.
Meta calls this Active Focus.
In practice, that means Muse Wonder is trained to:
Meta evaluates this behavior with FocusBench, an internal benchmark covering personal-agent tasks across domains such as housing, travel, IT, and finance.
For software work, sustained execution is also tested through SWE-Marathon, which evaluates long-running engineering projects.
The important distinction is that Active Focus is not just "more reasoning tokens." It is a behavioral goal: staying oriented toward the actual task over a long trajectory.
Software engineering brings memory and focus together.
A coding agent has to remember requirements, understand a codebase, make changes across multiple files, run tests, inspect failures, and continue iterating without losing the original objective.
Meta says both Muse Wonder models were trained on diverse, long-running software tasks for exactly this reason.
The strongest published coding results include:
The DeepSWE and SWE-Marathon results are particularly important because they focus on longer-horizon software engineering rather than isolated code completion.
Meta's official model page also describes Muse Wonder as able to plan long tasks, delegate work to subagents, call tools in parallel, and recover from errors.
The company demonstrates this broader coding capability with a playable 3D kitchen game generated from customizable image inputs.
That example is visually impressive, but the model card adds a useful limitation: precise visual and spatial reasoning, especially 3D tasks, remains a known weakness. Developers building applications that depend on exact spatial understanding should validate the model carefully.
A strong benchmark score does not always mean a model performs equally well on unfamiliar tasks.
Describe your idea once, and We0 AI can generate a showcase site, pages, and CMS, then help you attract customers and traffic after launch.
One complete project generation for free registration
Best for trying one complete generation flow and seeing a first project draft quickly.
Meta uses the term "benchmaxxing" to describe the concern that a model can be optimized heavily for known evaluations without developing the broader capability needed to handle new situations.
To study this, Meta compares performance on two groups of tasks:
The company compares Muse Spark 1.3 with Muse Wonder Full at multiple reasoning-effort levels.
According to Meta, both models improve on unfamiliar tasks as test-time compute increases. But Wonder Full performs better on out-of-distribution tasks than the older Spark trend would predict based only on its in-distribution performance.
Meta attributes that shift to a stronger underlying base model.
The practical argument is straightforward: a personal agent cannot be trained for every future combination of goals, tools, applications, constraints, and unexpected events.
It therefore needs to generalize beyond the tasks explicitly emphasized during post-training.
Meta demonstrates this idea with several personal-agent scenarios.
Muse Wonder can work in a browser and across applications to build a running plan using personal preferences and current circumstances.
The example combines personal context with external map and web information rather than treating route planning as a single isolated query.
Another demonstration uses real photos from a personal Instagram profile to design and order customized merchandise.
This is the kind of workflow where memory, multimodal understanding, browser use, and action-taking all need to work together.
Muse Wonder is also shown preparing a seating chart from a complex list of matching and placement requirements.
This is a useful example of constraint-heavy planning: the model must preserve many relationships and preferences at once while generating a final arrangement.
The original Meta page presents these scenarios as interactive videos rather than static screenshots. Because the demo media does not expose stable standalone image URLs suitable for Markdown embedding, those interactive examples are described here without inventing replacement images.
More capable agents also create a harder safety problem.
A personal agent may have access to:
That means safety is not only about whether the model answers a dangerous question. The system must also handle private information carefully and resist malicious instructions that may appear inside the content it reads.
Meta says Muse Wonder is its most aligned model family to date and that the models are trained for both privacy-sensitive behavior and adversarial robustness.
Meta evaluates Wonder Full—the more capable member of the family—under its Meta Superintelligence Scaling Framework.
The framework focuses on three catastrophic-risk domains:
The Muse Wonder model card says the unmitigated model is assessed at moderate or lower in cybersecurity and loss-of-control risk.
For chemical and biological risk, Meta says it does not rule out the unmitigated model reaching the high-risk threshold. Because of that, the company applies additional safety post-training and system-level classifiers.
With mitigations in place, Meta says the residual risk is assessed at moderate or lower across all three domains.
The model card reports the following capability results for Muse Wonder Full:
| Cyber Evaluation | Muse Wonder Full |
|---|---|
| CyberGym | 88.4% |
| ExploitBench | 58.9% |
| ExploitGym | 92 flags captured* |
The ExploitGym number needs an important footnote.
Meta says only 21 of the 92 captured flags used the intended vulnerability. The other 71 were obtained through unintended vulnerabilities. Because the published ExploitGym leaderboard counts intended-vulnerability exploitations, 21 is the comparable number for ranking purposes.
This is a good example of why benchmark methodology matters as much as the headline score.
Meta's broader Muse architecture also adds system-level protections around the model.
Muse runs agent workloads inside isolated infrastructure, keeps connected-service credentials separate from the agent environment, and routes external interactions through additional security controls.
Meta recommends human confirmation for irreversible actions and says developers should add guardrails appropriate to their own application.
The Muse Wonder model card also explicitly notes that:
Those limitations matter because an agent that can act is more consequential than a model that only produces text.
Muse Wonder is a frontier AI model from Meta designed for personal and professional agentic tasks. It focuses on long-running workflows, coding, computer use, Active Memory, and Active Focus.
Muse Wonder is available through the Meta Model API and Muse Code, while Muse Wonder Full is the higher-capability version offered through Muse Code subscriptions. Wonder Full leads more of Meta's published agentic and coding benchmarks, but the standard Wonder is the model directly available by API.
Meta's model card lists a 1-million-token context window for both Muse Wonder and Muse Wonder Full. The models accept text, image, video, and audio input and produce text output.
Meta currently lists Muse Wonder at $4 per million input tokens, $0.20 per million cached-input tokens, and $20 per million output tokens. Muse Wonder Full is not listed with per-token API pricing because it is currently distributed through Muse Code subscription plans.
Active Memory is Meta's term for carrying important user context across conversations and retrieving it when it becomes relevant. It is intended to make personal and professional agents more consistent over long-running interactions.
Active Focus is the model's ability to continue working toward a difficult goal, try alternatives, recover from blockers, and adapt as new information appears. It also includes asking for clarification and checking with the user before important actions.
Not universally. Meta's scorecard shows Muse Wonder models leading on some benchmarks while Claude Opus 5.5 and GPT-6 Astra lead on others, and the comparison mixes internal, third-party, and provider-reported results.
Meta's release describes Muse Wonder as available through the Meta Model API and Muse Code rather than as an open-weight release. Developers should use Meta's official model page for current availability and licensing terms.
Muse Wonder and Muse Wonder Full are Meta's latest frontier models for personal and professional agents. Their main design themes are Active Memory, Active Focus, domain expertise, long-running coding, computer use, and better generalization beyond the tasks emphasized during post-training.
The models also show why agent benchmarks need careful interpretation. Meta's published results combine internal, third-party, and provider-reported evaluations, and the model card is explicit about differences in harnesses and runtime settings. The family is highly competitive, but no single model wins every benchmark.
The larger change is architectural at the product level: Meta is trying to build agents that remember what matters, stay focused over long trajectories, and safely take action across real applications.
Muse Wonder's core idea is not simply a stronger chatbot—it is a model designed to keep context, keep working, and act as a persistent personal or professional agent.
Start from one sentence and have a complete website in minutes.
