Meta's Superintelligence Lab unveils its first public model, Muse Spark

Meta's newly formed Superintelligence Lab has unveiled Muse Spark, its first public AI model and a strategic pivot from the company's previous open-source Llama efforts. Unlike Llama, Spark is a proprietary model designed to leverage Meta's unique data advantage, integrating real-time content from platforms like Instagram, Facebook, and Threads to provide answers with citations and links to public posts. The model introduces a novel 'Contemplation' mode that orchestrates up to 16 reasoning agents working in parallel, which Meta claims delivers superior performance—including a top score of 58.4 on the challenging Humanity's Last Exam benchmark—without increasing latency.

Technically, Muse Spark represents a ground-up overhaul, emphasizing improved reliability through advanced reinforcement learning (RL) with 'thinking time penalties.' This system balances correctness with token efficiency, leading to a 'phase transition' where the model learned to compress accurate reasoning into significantly fewer tokens before scaling up again for higher accuracy. While Meta acknowledges gaps in areas like long-horizon agentic systems and coding, Spark is positioned as a competitor to models from OpenAI, Anthropic, and xAI. It launches with an updated AI safety framework and is available immediately via the Meta AI app and website, with a broader rollout across WhatsApp, Instagram, and Messenger planned for the coming weeks.