h3-rs
MiniMax H3 text/keyframe/reference-to-video-with-audio on your own desktop GPU, in Rust. One set of fused kernels runs through WebGPU (Vulkan, DX12, Metal) or CUDA; the CPU reference is the ground truth every kernel is checked against. Every recipe of the FastVideo MiniMax H3 cookbook is a --variant.
Variants
--variant | What it is | Forwards |
|---|---|---|
| h3 | MiniMaxAI/MiniMax-H3 base: T2VA, FL2VA, Ref2VA (dense attention) | 49 |
| fasth3-v1 | FastH3 4-step preview (VSA 0.9, tile 64) | 4 |
| fasth3-v2 | FastH3 8-step V2 (VSA 0.8, video shift 10) | 8 |
| fasth3-trim | FastH3 Trim: 42 blocks, rank-16 AdaLN, NVFP4 linears, 26-block decoder | 8 |
| fasth3-omni | FastH3 Omni Ref2VA with Parallel Decoding Distillation (32 heads) | 8 |
| pika-flex | Pika H3 Flex: FL2VA / Ref2VA LoRA, renoise sampler, 480p / 768p buckets | 6 |
| regen-2k | Open-H3-Regenerate-2K: 768p draft → 2560×1440 (h3 regen2k) | 3 |
Run the engine
Download the base model once (VAEs, text encoder, tokenizer) and the variant's DiT, then start the engine and allow this site to reach it:
huggingface-cli download MiniMaxAI/MiniMax-H3 --include "vae/*" "audio_vae/*" "text_encoder/*" "tokenizer/*" --local-dir h3
huggingface-cli download FastVideo/FastVideo-FastH3-8-Step-V2 --local-dir fasth3-v2
h3 adaln-tables --variant fasth3-v2 --model-dir h3 --variant-dir fasth3-v2 --out v2-tables.safetensors
h3 serve --variant fasth3-v2 --model-dir h3 --variant-dir fasth3-v2 --tables v2-tables.safetensors \
--weights int4-hqq-g64 --allow-origin https://h3-video.gialabs.sh
The engine listens only on 127.0.0.1 and answers this one site. Without --allow-origin it serves the studio itself at http://127.0.0.1:7860/.
Where it stands
- Every op and every model has a CPU reference;
h3 check --backend wgpu|cudacompares the GPU kernels against it. - No speed claim is made without a measurement on named hardware: see
docs/runs/in the repository. - Parity with the PyTorch reference is checked by fixtures generated from FastVideo / Diffusers / transformers (
scripts/reference/make_fixtures.py).