QuantFunc INT4 MiniMax H3: 3.2 s per step on an RTX 4090

The QuantFunc 4-bit build of MiniMax H3 runs at 3.2 seconds per sampling step on an RTX 4090 for a 768x768, 124-frame clip, which its card says is 3.19 times faster than an FP8 baseline at 10.2 seconds. That is a per-step number, not a per-clip time, and the card does not state a VRAM requirement, so do not size a purchase from it.

Everything about QuantFunc below is from its model card, read 2026-10-03, where it is among the models trending on Hugging Face's text-to-video list. Sume's side is from the video docs.

What does the QuantFunc build claim?

The card describes a 4-bit quantized MiniMax H3 that generates video with audio from text, images or video references. It says it runs on every NVIDIA GPU of compute capability SM75 or newer: the RTX 20, 30, 40 and 50 series, A100, H100, H200, B100, B200 and GB300. The speed claim is a measurement on one card: on an RTX 4090 at 768x768 and 124 frames, INT4 takes 3.2 seconds per step against 10.2 seconds for the FP8 baseline.

The card also reports quality against the BF16 original: about 23.7 dB PSNR, and a maximum relative output error of about 7.58e-4 from FP16 folding. PSNR of that size says outputs are close but not identical to BF16, so do not expect the same frame for the same seed.

How long is a clip, if a step takes 3.2 seconds?

The card gives no whole-clip time, so any figure is arithmetic on its per-step number and leaves out the text encoder, the audio path, the decode and model loading. As an illustration only: at 3.2 seconds per step, 8 steps is 25.6 seconds and 20 steps is 64 seconds of sampling at 768x768 and 124 frames on that card. Longer clips and higher resolutions cost more per step; the card does not say how much.

Measure end to end on your card and your workflow before you compare it with a hosted job. MiniMax H3's ComfyUI route against a hosted API lists the limits that change between the two.

Where does a hosted job still win?

Three places. Setup: the card says QuantFunc needs its own ComfyUI node or engine and does not work with standard Diffusers, which is one more runtime to maintain. Licence: the MiniMax H3 Community License covers the weights and derivatives, so the revenue and territory conditions still apply to your output. Scale: a card renders one clip at a time, while a hosted job queue takes many.

In Sume's Video generation docs, minimax-h3 accepts 5 to 15 seconds at native 480p or 768p, and each model lists pricing_skus in GET /v1/videos/models. Read the live price there rather than from a blog post, and compare it with your own cost per GPU hour divided by clips per hour.

Should you try the 4-bit build?

Try it if you already own an NVIDIA card in the supported range and want local iteration, and run a fixed prompt through the 4-bit build and your current precision to see the difference yourself. Skip it if you need exact reproducibility against BF16 outputs, or if you deliver a few clips a week, where a hosted job is cheaper than the hours you spend on setup.

Keep a record of the build and the settings you used, because a quantized checkpoint changes output compared with the original model and you will want to explain a difference later.