Meta's New Frontier: Introducing Muse Wonder and Muse Wonder Full
Meta has officially introduced Muse Wonder and Muse Wonder Full, its most advanced artificial intelligence models to date. These powerful models are engineered to drive advanced AI agents for both personal and professional applications, specifically targeting demanding, long-duration tasks across fields like coding, legal analysis, finance, and general computer use. Muse Wonder is immediately accessible via the Meta Model API and Muse Code, while the more powerful Muse Wonder Full is reserved for subscribers of paid Muse Code plans.
Advanced Capabilities and Core Features
Both models boast a substantial 1-million-token context window, facilitating extensive information processing. They support diverse input modalities, including text, images, video, and audio, with text output capabilities. Operating exclusively on Meta's cloud, access is provided through the Model API or the Muse Code command-line interface, compatible with macOS, Linux, and Windows.
Meta highlights enhancements from its Prometheus data center training, crediting a larger base model for improved performance on novel tasks. Key defining behaviors include:
- Active Memory: The model retains user context across interactions, retrieving relevant details as needed.
- Active Focus: It demonstrates persistence on challenging tasks, explores alternatives, and seeks confirmation before critical actions.
For developers, Muse Wonder is designed to break down large tasks, delegate to sub-agents, and interact seamlessly with browsers, desktop, and mobile applications.
Benchmark Performance: A Mixed Landscape
Evaluations indicate Muse Wonder Full's competitive standing, leading six out of fourteen categories in Meta's benchmark table. However, it trails Claude Opus 5.5 in several other tests. Notably, Muse Wonder achieved an 85.6 on MemoryBench, an internal Meta evaluation, marking the highest result among five compared models. Conversely, its ProgramBench score of 50.0 lagged significantly behind Claude Opus 5.5’s 70.5.
Key performance points:
- Muse Wonder Full excelled in DeepSWE 1.1 (79.0) and SWE-Marathon (57.5).
- Claude Opus 5.5 maintained superiority in ProgramBench, OSWorld, and Terminal-Bench.
- Meta notes competitor runs for internal benchmarks were "best-effort" and might not be fully optimized.
Access, Pricing, and Developer Integration
Muse Wonder is available at $4 per million input tokens and $20 per million output tokens, with a $0.20 cached rate. Muse Wonder Full requires a Muse Code subscription, starting from the Power plan ($50/month) up to the Maximum plan ($200/month).
Developers can initiate use by visiting dev.meta.ai, signing in, and generating an API key after adding billing information. The API supports OpenAI compatibility, allowing integration with the OpenAI SDK by specifying Meta’s base URL. Alternatively, Muse Code can be installed as a CLI tool. After installation and sign-in, users can select muse-wonder-1.0 to begin development.
Key Takeaways
- Meta's Muse Wonder and Muse Wonder Full are powerful new models for advanced AI agent work with a 1M context window.
- Muse Wonder Full shows strong performance in specialized coding tasks but faces stiff competition from Claude Opus 5.5 in other areas.
- Multi-modal input (text, image, video, audio) is a distinguishing feature.
- Pricing varies, with Muse Wonder available on a token-usage basis and Muse Wonder Full requiring a Muse Code subscription.
- Details regarding model parameters, open weights, or comprehensive safety reports are currently undisclosed.
This article is a rewritten summary based on publicly available reporting. For the original story, visit the source.
Source: MarkTechPost
