{
  "id": 14058276,
  "title": "Meta releases Muse Wonder AI models built to run personal and professional agents",
  "url": "https://urgent.news/2026/10/12/meta-releases-muse-wonder-ai-models-built-to-run-personal-and",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-12T22:36:05.000Z",
  "source": {
    "name": "SiliconANGLE",
    "slug": "siliconangle",
    "url": "https://siliconangle.com/2026/10/12/meta-releases-muse-wonder-ai-models-built-to-run-personal-and-professional-agents/"
  },
  "original_language": "en",
  "account": "Meta has unveiled Muse Wonder and Muse Wonder Full, its most advanced artificial intelligence models designed to manage personal and professional tasks effectively. The company hopes these models can handle longer, more complex jobs that people typically delegate to AI agents. One example given by Meta is processing an insurance claim. The new models boast strengths in coding and specialized fields such as law, finance, and healthcare.\n\nTwo crucial features, Active Memory and Active Focus, allow Muse Wonder Full to maintain context across conversations and remain productive even when tasks become challenging. If the model encounters difficulties, it will try alternative approaches before tackling the issue. Both models were trained at Meta's Prometheus data center, expanding on the pretraining and reinforcement learning methods introduced with Muse Spark in April. These models can process prompts mixing text with video, audio, and images up to a massive 1 million-token context window, with responses only providing text-based answers.\n\nComparatively, Meta's Muse Wonder Full tested against OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1 and Claude Opus 5.5. On the Artificial Analysis' Legal Agent Benchmark, Muse Wonder Full scored 13.1%, while GPT-6 Astra achieved the highest score at 8.6%. Muse Wonder Full showed better performance on coding tasks, leading in two long-horizon tests, DeepSWE 1.1 and SWE-Marathon, and scoring 57.5% on SWE-Marathon, compared to 46.3% for Claude Opus 5.5.\n\nIn terms of safety, Meta has yet to release a full safety report. However, the model card states that Muse Wonder Full has a \"moderate or lower\" risk rating for chemical and biological threats, including cybersecurity and loss of control, when operating under mitigations. ExploitGym tests revealed that 21 out of 92 captured flags came from unintended weaknesses, with the remaining 71 targeting intentional vulnerabilities.",
  "summary": "Meta Platforms Inc. today released Muse Wonder and Muse Wonder Full, two artificial intelligence models it says are its most capable yet at running agents that handle personal and professional tasks. Meta aimed the release at the longer, multistep jobs people hand off to an AI agent. One example it gave was working through an […] The post Meta releases Muse Wonder AI models built to run personal…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 2,
    "also_reported_by": [
      {
        "outlet": "Hacker News",
        "title": "Muse DAU dropped and Meta insiders decided to sell",
        "url": "https://urgent.news/2026/10/12/muse-dau-dropped-and-meta-insiders-decided-to-sell",
        "published": "2026-10-12T20:05:00.000Z"
      }
    ]
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}