Meta Muse Spark: First Model from Meta Superintelligence Labs
Meta launched Muse Spark, the first model from Meta Superintelligence Labs—the unit formed after the $14.3B Scale AI acquisition. The model introduces "Contemplating mode," an inference-time architecture running multiple agents in parallel then aggregating results rather than reasoning sequentially. It scores 50.2 on Humanity's Last Exam With Tools (vs Gemini 3.1's 43.9) and 38.3 on FrontierScience Research (vs GPT-5.4 Pro at 36.7). Notably, Meta abandoned its Llama open-source strategy for Muse Spark, which ships as a proprietary, login-gated model.
Why It Matters
Meta is making a deliberate break from its open-source Llama lineage and betting on inference-time multi-agent scaling as the differentiator—a direct challenge to OpenAI's and Anthropic's reasoning model strategies, and a signal that the AGI labs race has expanded to include Meta in earnest.