Nex-N2.5-mini · repaired
68,508 completion tokens · 2,745.2s wall time
Nex vs Ling: The Repair Is Part of the Story.
The same pagoda prompt on different deployments. Nex’s generated scene needed two missing declarations before it rendered; the untouched original and exact diff remain available.
Nex-N2.5-mini · BF16 · 1× Spark · repairedLing-3.0 Flash VL · FP8 · 2× Spark · original
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.
Ling-3.0 Flash VL · original
22,139 completion tokens · 862.2s wall time
The setup, the repairs, the limits.
Nex: one DGX Spark, BF16, eager execution, no speculation, 131,072 configured context, temperature 0.6 and no request output-token cap. Requested xhigh falls through to normal thinking in its native template; it is not a distinct supported effort.
Ling is the earlier two-Spark FP8 run: temperature 0.7, thinking enabled, xhigh requested and a 120,000-token cap. Different hardware, precision, sampling and cap policy make this a visual comparison—not a controlled speed contest.
The Nex repair adds only const lampSprites = []; and let waterMesh = null;. The saved verification exercised Day, Dusk, Night, Rotate, Petals and Reset. This is a ten-second render excerpt. Full-benchmark scores are separate and are not ranked here.
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.
The Ling output is the same saved generation used in both the Qwen vs Ling and Nex vs Ling comparisons, not an independent second sample.