{
  "id": 5195896,
  "title": "Zuck's Muse to Spark joy with open weights release 'soon'",
  "url": "https://urgent.news/2026/09/02/zucks-muse-to-spark-joy-with-open-weights-release-soon-5195896",
  "topic": "science",
  "section": "Science",
  "published": "2026-09-02T22:31:10.000Z",
  "source": {
    "name": "The Register Science",
    "slug": "the-register-science",
    "url": "https://www.theregister.com/ai-and-ml/2026/09/02/zucks-muse-to-spark-joy-with-open-weights-release-soon/5294093"
  },
  "original_language": "en",
  "account": "Meta, led by CEO Mark Zuckerberg, is set to unveil an open weights version of its leading AI model, Muse Spark, in the near future. In the meantime, Muse Spark 1.3, a smarter, more efficient iteration of the model, is already available on Meta's API service and Muse Code CLI. Since its April release, multiple refinements have been rolled out to maintain pressure on competitors and demonstrate the value of Meta's significant investments.\n\nMuse Spark 1.3 brings significant improvements, particularly for users working with AI agents and code assistants. This update is particularly welcome, considering the release of OpenClaw 2.0. According to Meta, the model should perform better in long-term tasks and be more realistic about its capabilities, avoiding repetitive dead-end paths. It will seek human input more often when the correct course is unclear, ask clarifying questions for ambiguous prompts, and confirm before undertaking consequential actions.\n\nThese enhancements not only make the model more reliable but also help reduce the number of turns and tokens needed to complete a task, potentially lowering costs. Compared to Muse 1.2, Meta claims Muse Spark 1.3 delivers solid gains across a wide range of benchmarks. Independent testing by Artificial Analysis suggests Spark 1.2 is already competitive, performing on par with GPT 5.6 Terra and Z.AI's GLM 5.3 Flash. The latest benchmark results largely support Meta's claims, showing a 4-point increase in overall intelligence, putting it in a tight race with GPT 5.6 Sol, Claude Opus 5, and Grok 4.6 High.\n\nFor those who prefer lower rate limits or are willing to grant Meta access to their prompts for training data, Meta offers contributor tier discounts, starting at $0.002 per million tokens for cached input, $0.10 per million tokens for input, and $0.20 per million tokens for output.",
  "summary": "While you wait, Meta says it’s taught the model to stop wasting tokens and ask for help a bit more often",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 2,
    "also_reported_by": [
      {
        "outlet": "The Register",
        "title": "Zuck's Muse to Spark joy with open weights release 'soon'",
        "url": "https://urgent.news/2026/09/02/zucks-muse-to-spark-joy-with-open-weights-release-soon",
        "published": "2026-09-02T22:31:10.000Z"
      }
    ]
  },
  "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."
}