Qwen Flash-Next FP8 · TP2
52,046 completion tokens · 1,292.8s wall time · finish: stop
Fire in a Glass Box
A spark, growing fire, and a wind-driven blowout. Three deployments take the same creative brief from reasoning to an animated installation.
Qwen Flash-Next · FP8 · 2× SparkQwen 27B AEON · NVFP4 · 1× SparkQwen 27B · MLX 8-bit · Mac Studio
Left-to-right panel order matches the labels above. These are browser recordings of generated code, not recordings of model inference. On mobile, rotate your device or use fullscreen to inspect details.
Qwen 27B AEON NVFP4 · TP1
27,498 completion tokens · 1,398.2s wall time · finish: stop
Qwen 27B MLX 8-bit · Mac
85,305 completion tokens · 3,701.1s wall time · finish: stop
The setup, the repairs, the limits.
Thinking enabled; xhigh was requested; temperature 0.7; explicit 120,000-token request budget. These historical runs were capped, not uncapped.
Qwen Flash-Next required a hex-color literal repair. AEON required a library-URL correction and script-tag cleanup. The Mac artifact has no recorded repair. Generated originals are linked beside repaired versions.
The page shows a scene-only crop of the existing comparison recording. Historical stat overlays were removed; per-arm usage and wall time below come from the saved generation metadata. These are deployment-level comparisons, not quant-only controls.
Completion counts include reasoning. Reasoning character counts are not token counts. Saved wall times use each run’s original measurement boundary; no cross-model speedup or causal quantization claim is made here.