Meta launches Muse Spark from Superintelligence Labs — Alexandr Wang's first proprietary flagship targets multimodal reasoning at sub-Llama-4 compute cost

Meta unveiled Muse Spark as the first flagship model from the newly-formed Superintelligence Labs unit under Chief AI Officer Alexandr Wang. The proprietary model competes on multimodal perception, reasoning, health, and agentic-task benchmarks at a fraction of the compute cost of Llama 4's mid-size variant — signaling Meta's strategic pivot from open-source-only to a proprietary + open-source two-track model program.

Muse Spark is the explicit signal that the Llama-only era is over. Wang took over AI strategy after the Scale AI acquisition; Muse Spark is the first model program shipped under his governance. The performance-per-compute claim is the key marketing axis — Meta is positioning Muse Spark as a model that wins on TCO for enterprise inference rather than on absolute capability benchmark leadership. That maps to Meta's recently-announced $115-135 billion 2026 AI capex (nearly double 2025 spending) — the strategic premise is that inference economics, not training-flop budget, determines the long-run winner.

The open-source-or-not status of Muse Spark itself is not yet public. Meta's two-track positioning — Llama as open-source community anchor, Muse Spark as proprietary enterprise flagship — would let it sell against both Anthropic's Claude (proprietary-only) and the open-weight Chinese lab releases.

Meta AI Blog — Muse Spark — Superintelligence Labs first flagship → · Reuters — Meta launches Muse Spark from Superintelligence Labs →