Meta Launches Muse Spark From New AI Lab

Meta released Muse Spark, the first model out of its newly formed Meta Superintelligence Labs, built as a natively multimodal reasoning system with tool use, visual chain of thought, and multi-agent orchestration baked in from the start. The pretraining stack behind it required over an order of magnitude less compute than Llama 4 Maverick to reach equivalent capabilities - a significant efficiency claim if it holds up externally. A Contemplating Mode runs parallel reasoning agents and scored 58% on Humanity's Last Exam. Health reasoning was trained with input from over a thousand physicians. Meta is clearly trying to separate this line from the Llama open-weight track, positioning MSL as its frontier lab in the same mold as OpenAI's o-series and Anthropic's Claude reasoning models.