Meta Muse: Spark 1.2 vs Glimmer 30B
Meta shipped two new Muse models, so we ran both through the same 35 one-shot prompts. Neither failed to ship a single build — 35 of 35 runnable each — so the gap between them is entirely about quality, not reliability.
Muse Spark 1.2, the million-token flagship, comes out ahead at 3.38 average, with 21 of 35 builds at four stars or better and the pair's only clean five: a flow-field particle system with fading trails and reactive input. Glimmer-30B trails at 2.89. It's steady on simulations — boids, the chaotic double pendulum, Lorenz, a fully playable 2048 — but brittle on canvas games, where six builds crash straight to a black screen (a black alien-shooter, an asteroids that dies on Start, a fluid-simulation that never paints). Spark's misses are milder and fewer: a frozen Mario, a maze that never draws, a Rubik's cube that renders only its UI overlay.
Same family, same clean build rate — Spark just holds together on the hard rendering prompts where Glimmer cracks. The head-to-head prompts are below; each build links to its live artifact and the grade behind it.
