Meta Launches Muse Spark, Its First Proprietary AI Model From Superintelligence Labs

Meta has released Muse Spark, a new proprietary AI model developed by its Meta Superintelligence Labs division, marking a significant strategic shift away from the company's open-source Llama family. The model, led by Chief AI Officer Alexandr Wang, formerly of Scale AI, is available initially through the Meta AI app and website, plus a private API preview for select users. Unlike the Llama models, Muse Spark will not be released as open weights. Muse Spark features what Meta calls "visual chain of thought" capabilities, allowing the model to annotate dynamic environments and reason through visual information. The system also includes a "Contemplating" mode that orchestrates multiple sub-agents to reason in parallel. Benchmark results show Muse Spark achieving a score of 52 on the Artificial Analysis Intelligence Index v4.0, a significant improvement from Llama 4 Maverick's score of 18. The model scored 80.4 on MMMU Pro for vision tasks and 42.8 on HealthBench Hard, outperforming competitors including GPT-5.4 and Claude Opus 4.6 in medical reasoning. Meta reports that Muse Spark achieves these results using over an order of magnitude less compute than its predecessor, attributed to a technique called "thought compression" that penalizes excessive reasoning time during training. The launch raises questions about the future of the Llama model family, which has been widely used by developers and researchers since its introduction in 2023.