Summary
Meta’s Muse Spark is the first model to emerge from Meta Superintelligence Labs, the division built after Llama 4 underdelivered and CEO Mark Zuckerberg made the decision to tear down and rebuild the company’s entire AI stack from scratch. Led by Alexander Wong, brought in from Scale AI, the new team spent nine months on a new architecture, new data pipelines, and new training infrastructure. The result is a natively multimodal, reasoning-capable model now powering Meta AI across WhatsApp, Instagram, and Facebook — reaching over 3 billion users daily.
Fahd Mirza runs Muse Spark through four practical tests via meta.ai: correlating five simultaneous medical images (chest X-rays, microscopy slides, and a blood report), generating a fully playable retro space shooter game in a single HTML file, and identifying two visually near-identical flags (Indonesia and Poland) with historical context. The model handles all four with notable competence, including appropriate epistemic humility on the medical task. Benchmark comparisons shown in the video place Muse Spark ahead of several frontier models on HealthBench (42.8 vs. Gemini 3.1 Pro at 20 and Opus 4.6 at 14), though results are more mixed on other dimensions.
The video offers a grounded first look at a model that represents a significant organizational and architectural bet by Meta. For those tracking the competitive frontier model landscape, Muse Spark’s multimodal reasoning capabilities and the story behind its development make it worth examining closely.
📺 Source: Fahd Mirza · Published April 09, 2026
🏷️ Format: Review
