Muse Spark drops with Contemplating Mode

Muse Spark marks Meta's definitive pivot from open-source Llama to proprietary models. Its Contemplating Mode uses a parallel multi-agent approach to rival frontier reasoning models without high latency. Native multimodality enables unified perception and visual chain-of-thought, while health benchmarks suggest a focus on specialized verticals. Claimed 10x compute efficiency over Llama 4 indicates a major architectural leap, signaling a shift toward subscription monetization.