Meta unveiled Muse Spark on Wednesday, the inaugural AI model from its Superintelligence Labs and the first in the Muse family. The company described it as a ground-up overhaul of its AI efforts, aiming for personal superintelligence. While proprietary for now, future open-source models are planned.
Meta announced Muse Spark on April 8, marking a departure from its previous Llama models. Formed less than a year ago, Superintelligence Labs released this model with the goal of delivering personal superintelligence to everyone. Meta founder Mark Zuckerberg stated on Threads that the Muse family will include new open-source models in the future, though Spark remains proprietary initially. The model draws on content from Instagram, Facebook, and Threads, currently linking to public posts on locations or trends, with plans to integrate Reels, photos, and recommendations while crediting creators. Meta highlighted strong benchmark performance in standard thinking mode, comparable to or better than models from OpenAI, Anthropic, Google, and xAI. It also introduced 'Contemplating' mode, which orchestrates up to 16 agents reasoning in parallel for superior results with similar latency, achieving 58.4 on Humanity’s Last Exam using external tools. The company acknowledged ongoing investments needed in long-horizon agentic systems and coding workflows. Reinforcement learning enhancements, including 'thinking time penalties,' enabled smoother gains and more efficient reasoning, as seen in a 'phase transition' on the AIME 2025 benchmark where accurate outputs used fewer tokens. An updated Advanced AI Scaling Framework deems the model safe across measured frontier risks, with further details in an upcoming Safety & Preparedness Report. Muse Spark is available now via the Meta AI app and meta.ai website, plus a private preview API for select partners. Rollout to WhatsApp, Instagram, Facebook, Messenger, and AI glasses will follow in coming weeks.
