Muse Spark

First model from Meta Superintelligence Labs (led by Alexandr Wang) and the start of the Muse model family — a complete departure from the Llama lineage, rebuilt from scratch over a nine-month sprint (internally codenamed "Avocado"). Natively multimodal: processes text, images, and audio in a single architecture with built-in tool use, visual chain-of-thought reasoning, and multi-agent orchestration. 260K token context. Parameter count undisclosed.

AA Intelligence Index: 52 — #4 globally behind GPT-5.4 (57), Gemini 3.1 Pro (57), and Claude Opus 4.6 (53). Reaches Llama 4 Maverick's capabilities with 10× less compute and used just 58M output tokens to complete the full AA evaluation (vs 120M for GPT-5.4, 157M for Opus 4.6). Strongest on healthcare (HealthBench Hard: 42.8, beating GPT-5.4's 40.1) and chart comprehension (CharXiv: 86.4, #1 overall). Humanity's Last Exam: 58%, FrontierScience Research: 38% (Contemplating mode).

Not open-weight — Meta's first proprietary model, available only via meta.ai, the Meta AI app, and a private API preview. Marks Meta's strategic pivot toward a closed commercial track alongside the open Llama family.