Quantized H3 vs lossless H3 vs a hosted clip: compare fairly

Do not put a quantized ComfyUI clip and a lossless SGLang clip in the same benchmark row. MiniMax's self-hosting guide says exactly that: ComfyUI uses pruned INT8 or NVFP4 weights, SGLang uses the official mixed BF16/FP32 weights, and their outputs should not be mixed. A hosted clip is a third case, and on Sume it also cannot be matched by seed.

What is different between the three?

How do you compare them fairly?

Use the same prompt, the same first frame if any, the same duration, and the same aspect ratio. The local SGLang output is 1344 x 768 at 24 FPS; ask for 768p and 16:9 on Sume. Do not require identical pixels, since you cannot seed the hosted one. Instead, run several samples per route and judge them blind.

• Write down the file name and loader for each local clip.
• Use at least five prompts per route; one prompt tells you little.
• Score what you care about: lip-sync, text, motion, audio.
• Keep cost on a separate line from quality.

What should the scorecard contain?

One row per clip, with the route, the file or model id, the prompt, the duration, the resolution, the wall-clock time, and a short note on what you saw. Add a cost column last: for local runs, GPU-seconds times your rate; for the hosted run, usage.cost from the finished job, which on Sume is the billable amount.

Then look for patterns, not winners from one clip. If the quantized file loses on faces but ties on motion, that tells you where to spend. If all three routes tie for your use case, the cheapest and simplest one wins.

What does the card promise?

The model card lists 24 FPS, 32 kHz stereo audio, durations of 4 to 15 seconds and 768p as the default. The hosted minimax-h3 row takes 5 to 15 seconds, so test a clip near 5 seconds on both. The hosted H3 can also upscale to 2K and 4K; the card lists its 2K regeneration module as API only. See the video docs.