MiniMax-H3-curve

About

What MiniMax-H3-curve is

MiniMax-H3 GGUF - curve form (fl2va + ref2va) GGUF quantizations of MiniMax-H3's 33B video+audio DiTs, built from MiniMax's pruned checkpoints. Same model, same quant tiers as the [original-form repo](https://huggingface.co/joeygambino/MiniMax-H3-GGUF), about 40% smaller - and a Q80 that fits a 24 GB card. fl2va = text/first-last-frame to video+audio (T2V and I2V). ref2va = reference-conditioned generation (identity from up to 9 images, 3 videos with soundtracks, 3 voice clips). Runs in ComfyUI with ComfyUI-GGUF plus a one-line architecture patch (node pack: ComfyUI-H3-Multishot; workflows: MiniMax-H3-Multishot-Workflow). Requires ComfyUI 0.30.0 or newer On anything older these will not load at all, because the shape of the modulation weights changed and older builds do not know how to read them. That is the only catch - everything else is a drop-in swap: same node, same workflow, same filenames apart from -curve-. If you are pinned to an older ComfyUI, or you would simply rather run the larger quants, the original-form files are still maintained at joeygambino/MiniMax-H3-GGUF. That repo is not going anywhere. Where the 40% went About 40% of this model was never unique data. Each block carried a 96768 x 2688 adaln modulation matrix, and MiniMax's pruned checkpoints show those matrices are a smooth function of the timestep rather than 51 independent tensors. The pruned form factorizes each one to 96768 x 8 and adds a single shared adalnttable to reconstruct them. Same math, ~26 GB less file. This is why Q80 exists here and not in the original repo. In the original form it was pointless - it landed at roughly the size of the official int8 release. Factored down, Q80 is smaller than the Q51 in that repo (21.5 GB vs 25.9 GB). Is it actually the same? Checked by rendering, not by assuming. Every comparison below is the same graph and the same seed, changing only the model file: Numbers are mean absolute pixel difference over all 124 frames. The noise floor being exactly 0.00 is what makes the rest meaningful - sampling is deterministic at a fixed seed, so these are real differences and not run-to-run variation. Swapping to curve form moves the output about a third as

Summarised from the published model card. Read the full card on the HuggingFace links below.

Specifications

What it is

How it works

How video models work

Variants

Sizes & precisions

Open weights ship in several sizes and precisions. One page, all the variants — pick the one that fits your GPU. VRAM figures are estimates from model size.

Using it via the API

Call it like any OpenAI endpoint

Once AxForge deploys minimax-h3-curve for you, it answers on the OpenAI-compatible API — the same base URL and keys as every other model. (minimax-h3-curve below is illustrative; you get the exact model name on deployment.)

$ curl -sS https://api.axforge.ai/v1/videos/generations \
  -H "Authorization: Bearer $AXFORGE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"minimax-h3-curve","prompt":"a drone shot over a forest"}'

Details

Languages, data & research

Licence

Commercial licence needed

Sources

Weights & code

Explore

More video models