Meta Debuts MuseSpark: First Proprietary Closed Model Near Top of Leaderboard
Meta has released MuseSpark, its first major proprietary closed-source model — a notable departure from the open Llama line that defined Meta AI's previous positioning. Estimated at approximately 70B base parameters with 16 parallel agents yielding over 1 trillion effective parameters, it places just below Claude on the leaderboard cited in early benchmarks. Alibaba's Qwen 3 Ultra is also in this trillion-parameter MoE tier.
Why It Matters
Meta choosing closed-source for its most capable release signals that even the company most associated with open AI models now views competitive frontier capability as too strategically valuable to open-source — a meaningful shift for the open-weights ecosystem.