Meta’s Muse Spark is here – and it’s closed source

Meta Superintelligence Labs has released Muse Spark, its first AI model built from scratch over nine months under chief AI officer Alexandr Wang. The natively multimodal model introduces a 'Contemplating' reasoning mode using parallel sub-agents and a 'thought compression' training technique that claims 10x compute efficiency over Llama 4 Maverick. Benchmarks show a mixed picture: Muse Spark ranks fourth on the Artificial Analysis Intelligence Index, trailing Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6 on abstract reasoning (ARC AGI 2: 42.5 vs rivals' ~76), but leads on HealthBench Hard (42.8%) and chart understanding. Notably, Muse Spark is closed source — a significant departure from Meta's Llama open-source heritage — which Meta frames as temporary. The model now powers Meta AI across the company's platforms, with dedicated shopping and health modes leveraging Meta's platform data advantages. The release is the first product output from Meta's $14.3 billion Scale AI investment and its broader $115–135 billion 2026 capex plan.