Qwen Flash-Next FP8 · TP2
65,985 completion tokens · 1,664.1s wall time · finish: stop
A World Built. A Failure Kept.
Two generated floating islands, plus the missing third result. The AEON arm did not produce a working artifact in either of its two valid model attempts.
Qwen Flash-Next · FP8 · 2× 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 MLX 8-bit · Mac
65,536 completion tokens · 2,744.2s wall time · finish: stop
The setup, the repairs, the limits.
Successful displayed runs requested thinking ON, xhigh, temperature 0.7 and a 120,000-token output budget. Different model variants and engines mean this is not a controlled quantization-only experiment.
AEON attempts 1–2 failed because of harness limits (an output cap and a client timeout). Attempts 3–4 are distinct model failures: repetitive degeneration, then a natural stop with no final content. These must not be counted as four model failures.
The Mac scene includes a two-line duplicate-identifier repair. Flash-Next has no recorded repair. This excerpt preserves the original recordings’ motion; it is not a rendering-FPS benchmark.
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.