Summary
Sam Witteveen delivers one of the more substantive early analyses of Meta’s Muse Spark, framing the release not just as a model launch but as the first tangible output of a $14 billion organizational bet. After Llama 4 landed with a thud a year ago, Mark Zuckerberg dissolved the existing Meta AI structure and rebuilt it around a new division called Meta Superintelligence Labs, led by Alexander Wang recruited from Scale AI. What followed was a months-long hiring spree — with reported pay packages reaching hundreds of millions in equity — aimed at pulling top researchers from OpenAI, DeepMind, and elsewhere.
Witteveen is precise about what Muse Spark is and isn’t. Third-party benchmarker Artificial Analysis places it in the global top five, ahead of Claude Sonnet and several other strong models, and ranks it the second most capable vision model they’ve evaluated. The model comes in instant and thinking modes, with a rumored deep-think “contemplating” mode forthcoming. On the downside: no open weights, no public API, no technical transparency on training data or token counts, and weak performance on agentic benchmarks — an irony given that Meta acquired Manus partly to strengthen that capability. The model also finishes last on Humanities Last Exam with tools, a benchmark created by Alexander Wang’s own former company Scale AI.
The broader argument Witteveen makes — that money and talent acquisition have real but bounded returns at the frontier — gives this video weight beyond the immediate release cycle. For anyone tracking Meta’s trajectory from open-source leader to proprietary competitor, this is essential context.
📺 Source: Sam Witteveen · Published April 09, 2026
🏷️ Format: News Analysis
