Meta launches Muse Spark, its first closed frontier model
Led by former Scale AI chief Alexandr Wang, the model is proprietary and API-only, reversing the open-weight approach Meta had used for the Llama family.
- Models & capabilities
- Open weights & ecosystem
- Labs & people
- Major
Meta released Muse Spark, the first model from its recently formed Meta Superintelligence Labs, as a closed, proprietary system available only through the Meta AI app and a private API preview — not as downloadable open weights. Meta described it as a natively multimodal reasoning model supporting tool use, visual chain-of-thought and multi-agent orchestration, and said it required roughly an order of magnitude less compute to train than Llama 4 Maverick for comparable capability.
The release marked a reversal for a company that had built its public AI identity around openly released weights, from Llama 2 through Llama 4. Meta gave no statement directly addressing the change in strategy, and the closed release sat alongside Llama’s continued existence as a separate, still-open product line. Reporting on the shift pointed to three converging pressures: criticism that Llama 4’s benchmark reporting had been misleading, which undercut the reputational case for openness; the roughly $14 billion investment Meta made in Scale AI to bring in the company’s co-founder, Alexandr Wang, as chief AI officer, which built a proprietary data pipeline better protected by closed weights; and concern that competitors were using Meta’s open releases to train cheaper knockoff models through distillation.
Meta reported benchmark scores including 58% on Humanity’s Last Exam and 38% on a frontier-science research benchmark in the model’s “contemplating” mode, which runs parallel agent reasoning. Independent tracking placed Muse Spark fourth on the Artificial Analysis Intelligence Index at launch, behind Gemini, GPT and Claude’s leading models.
The launch was Meta’s first major model release under Wang’s leadership and was widely read as a test of whether the company could compete at the frontier without the goodwill its open releases had earned it. It also removed one of the few remaining Western labs contesting the top of the leaderboard with an open-weight model, leaving that role increasingly to Chinese labs such as DeepSeek and Alibaba.