MiniMax-H3-cmf

About

What MiniMax-H3-cmf is

MiniMax-H3 Turbo — one file, no Python MiniMax-H3 renders video and synchronized stereo audio from one prompt, in one transformer, on two flow schedules. larryvrh's Turbo LoRA brings it to four sampling steps. This is both of them in the CMF container — the DiT, the Qwen3-VL prompt encoder, the video VAE decoder and the audio vocoder in a single memory-mapped file — running on cortiq, a Rust binary with no ML framework underneath. The reference checkout is 124.4 GB across four files plus a ComfyUI install. Here it is one file between 13.2 and 23.9 GB, and which one you take is decided by your VRAM. Same prompt, three files "A corgi in a chef hat flipping a pancake, sizzling sounds and a cheerful bark." — 512×288, 39 frames, seed 42, four steps, nothing but the prompt. The 13.2 GB file is the one to take if you have 16–20 GB of VRAM: it holds the whole run resident instead of paging, and it is still four bits everywhere. Two bits is published as a dead end, not a choice — why, with numbers. The GIFs are silent and the audio is half the model, so take an mp4. One transformer denoises picture and sound together in one packed sequence. Which file Text-to-video everywhere; the fl2va files also take a first and/or last frame (--first-frame/--last-frame, binary P6 PPM). The compact clipproj4b-fl2va file was verified frame-for-frame: the conditioning image comes back as frame 0 of the render. One honest caveat: its ClipProj projection was fitted on text-only encoder activations, so with a picture in the prompt the conditioning quality on complex scenes is still being compared against the full-encoder fl2va files — report what you see in the discussions. (keyframes). The release's third path, ref2va, is not ported. The Turbo LoRA is merged into the weights, so the file IS the 4-step model — nothing else to download. 47.83 B parameters, 2 361 tensors, cortiq verify clean. Keyframe to video One picture conditions the run twice, and both halves matter. Its VAE latent becomes a row the DiT holds at a timestep of its own near 1 — a condition, not noise being removed — and never denoises. The picture ITSELF goes to the prompt encoder through Qwen3-VL's vision tower, as "<Pictu

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-cmf for you, it answers on the OpenAI-compatible API — the same base URL and keys as every other model. (minimax-h3-cmf 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-cmf","prompt":"a drone shot over a forest"}'

Details

Languages, data & research

Licence

Open weights

Sources

Weights & code

Explore

More video models