What we tested, exactly as it ran

Each report shows the task, every attempt, each check we ran, and the raw evidence — with speed reported separately. Nothing is reduced to a verdict; the numbers are what the runs produced.

01 · the checks

Every check we ran, shown with its result.

02 · every attempt

Cold, warm, and repair runs are all kept.

03 · measured speed

Timing recorded under one protocol.

checked runs·one entry per task protocol

task / model	protocol	checks held	status	runs	best measured rate	report
qwen3.6:35b release-schedule-v1	v1.0.0	11 / 12 91.7%	Measured	4	76.2 tok/s	open →
qwen3.6:35b advanced-creative-continuity-v1	v1.0.0	11 / 14 78.6%	Measured	4	73.2 tok/s	open →
qwen3-coder-next:q4_K_M defensive-zip-extractor-v1	v1.0.0	6 / 7 85.7%	Measured	1	43.3 tok/s	open →

These are single measured runs, not universal model rankings. A category becomes comparative only after two or more models run the same task and protocol.

observed runs·shown side by side, as measured

Four sampled frames of a red biped robot walking through a rainy neon alley

MiniMax H3 & Wan 2.2 — Local Video Generation

Two local video workflows ran on ComfyUI for Windows. Both clips are shown with their exact frames, resolution, timing, and audio state — MiniMax H3 followed the walking action with audio; Wan 2.2 held a coherent neon scene with a wheeled subject.

read observation →