Meta Platforms Inc. has released its latest artificial intelligence models, Muse Wonder and Muse Wonder Full, designed to handle complex personal and professional tasks. These models aim to improve AI agent capabilities for multistep jobs, building on Meta's previous work with AI. The announcement was first reported by Silicon Angle.
The new models, Muse Wonder and Muse Wonder Full, are Meta's most advanced yet for AI agents performing tasks such as managing insurance claims, coding, and specialist work in law, finance, and health. Key features include Active Memory, which maintains user context across conversations, and Active Focus, designed to help models overcome difficulties and recover from errors by trying alternative approaches. Prompts can incorporate text, video, audio, and images within a 1 million-token context window, with text-only outputs. Meta compared Muse Wonder Full against competitors like OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1 and Claude Opus 5.5, showing strong performance in legal and coding benchmarks. The company also addressed safety concerns, noting that while potential risks exist without mitigations, safeguards are in place to manage threats in areas like cybersecurity and chemical/biological risks. Muse Wonder is positioned as Meta's most aligned model to date, with a 0% harmful-action rate on agentic misalignment tests.
Source: Silicon Angle
