WaveCut/MiniMax-H3-OrbitQuant-W4A4

MiniMax H3 — OrbitQuant W4A4 with source FP32 VAEs

OrbitQuant conversion of MiniMaxAI/MiniMax-H3, pinned to source revision 73372e6cf53e414edd3ab03e357717fb0602e758.

Eligible linear weights in transformer, transformer_ref, and the Qwen3-VL text_encoder are stored and executed through OrbitQuant's native packed W4A4 path. Non-quantized boundaries use BF16 compute. The visual vae and audio_vae are byte-for-byte FP32 source copies and are never quantized.

Final 608×480 example

H.265 10-bit CRF 10 · H.264 fallback · CRF 1 yuv444p master · 16-frame overview · adjacent-frame review · audio spectrum

This live ComfyUI run uses 608×480, 124 frames at 24 FPS, seed 42, and 24 sigma points / 23 denoiser forwards. All 300 eligible denoiser linears use the native packed W4A4 path with no exact INT8 weight cache. Source FP32 tiled visual decode produced the retained CRF 1 master; the HEVC card copy was derived from that master at CRF 10. The output contains AAC stereo at 32 kHz.

Full-resolution frames and adjacent triplets were reviewed for face geometry, eyes, lips, grid artifacts, ghosting, texture breakup, and abrupt section redraw. The macro-to-face shot remains coherent. The audio spectrum is broadband without a persistent narrow electronic whistle.

ComfyUI workflow

Download the ready-to-import MiniMax H3 OrbitQuant T2VA workflow. It is based on Comfy-Org's bundled video_minimax_h3_t2v.json and preserves the official preset's readable composition.

The PNG above is a 3060×1310 ComfyUI Workflow Image Export, not a browser screenshot. Its tEXt workflow chunk contains the same six-node graph with balanced, 608×480, 124 frames, 24 steps, and the detailed example prompt.

Install ComfyUI-OrbitQuant into ComfyUI/custom_nodes, restart ComfyUI, import the workflow, and set OrbitQuant Release Loader.model_path to this downloaded model directory. The graph uses only the generic public nodes OrbitQuant Release Loader and OrbitQuant Generate Video; there are no MiniMax-specific public node classes.

On the RunPod ComfyUI image, launch ComfyUI with:

python main.py --listen 0.0.0.0 --port 8188 \
  --disable-cuda-malloc \
  --disable-dynamic-vram \
  --disable-async-offload

These supported flags let the OrbitQuant subprocess enforce its own allocator cap instead of competing with ComfyUI's global DynamicVRAM and async-offload layers.

Inference profiles

All numbers use CUDA 13, 608×480, 124 frames, 24 sigma points / 23 forwards, native-auto Torch Flash SDPA, no weight cache, sequential CUDA text conditioning, and source FP32 VAEs.

Profile	GPU	Task	Placement	Process peak	Denoise	Generation
balanced (default)	RTX PRO 6000	T2VA	streamed leaf offload, 12 GiB cap	6.36 GiB child; 6.90 GiB incl. idle ComfyUI	46.68 s	—
speed	RTX PRO 6000	T2VA	resident transformer	21.14 GiB	46.84 s	51.10 s
minimum_vram	RTX 4090	T2VA	low-CPU-memory streamed leaf offload, 8 GiB cap	4.07 GiB	154.25 s	188.70 s
speed	RTX PRO 6000	Ref2VA	resident transformer_ref	24.06 GiB	118.48 s	155.42 s

balanced is the recommended Pareto recipe. On the tested PRO 6000, streamed weight movement overlaps denoising closely enough to match the resident path while cutting the child process's physical CUDA peak by about 70%. minimum_vram is the verified absolute-minimum endpoint. speed removes transformer transfers when VRAM is available.

SageAttention2's available CUDA 13 binary did not include SM120 code for this PRO 6000, and forced cuDNN attention was slower. Native-auto Torch Flash SDPA is therefore the shipped supported attention path.

Install

pip install "orbitquant[hf,kernels]>=0.9.2,<0.10"
pip install "diffusers @ git+https://github.com/huggingface/diffusers.git@abc5e9bf71fd38f53cd471bc3acaa84bc5ecbfdc"
pip install "transformers>=5.13,<6" accelerate av soundfile

Or install all pinned runtime requirements from this repository:

pip install -r runtime-requirements.txt

Direct runner

The runner writes each scheduler checkpoint atomically and saves the latent bundle before decode. The examples below keep the prompt in a file to avoid shell quoting a multi-kilobyte description.

Balanced T2VA:

python scripts/run_quantized_example.py \
  --release . \
  --output balanced.mp4 \
  --save-latents balanced.latents.pt \
  --prompt "$(cat prompt.txt)" \
  --seed 42 --width 608 --height 480 --num-frames 124 --steps 24 \
  --manual-stage-offload \
  --text-encoder-sequential-offload \
  --transformer-group-offload-type leaf_level \
  --group-offload-use-stream \
  --cuda-memory-cap-gib 12 \
  --transformer-runtime-mode auto_fused \
  --checkpoint-dir checkpoints/balanced

Maximum-speed T2VA: remove the group-offload and allocator-cap options while keeping --manual-stage-offload --text-encoder-sequential-offload.

Minimum-VRAM T2VA: use the balanced command with --group-offload-low-cpu-mem-usage --cuda-memory-cap-gib 8.

Ref2VA speed:

python scripts/run_quantized_example.py \
  --release . \
  --output ref2va.mp4 \
  --save-latents ref2va.latents.pt \
  --prompt "$(cat prompt.txt)" \
  --task ref2va --reference reference.png \
  --seed 42 --width 608 --height 480 --num-frames 124 --steps 24 \
  --manual-stage-offload \
  --text-encoder-sequential-offload \
  --reference-vae-sequential-offload --reference-vae-tile-size 128 \
  --transformer-runtime-mode auto_fused \
  --checkpoint-dir checkpoints/ref2va

Decode only after the latent-producing process exits:

python scripts/decode_h3_latents.py \
  --latents balanced.latents.pt \
  --vae vae \
  --audio-vae audio_vae \
  --output balanced.master-crf1.mp4 \
  --preview-output balanced.mp4

The decoder always loads the release's untouched source FP32 visual and audio VAEs. The visual VAE is tiled and sequentially offloaded; the audio VAE enters GPU only for the audio stage.

Component precision and size

Component	Stored mode	Artifact GiB	Eligible linear coverage	OrbitQuant modules	AdaLN INT4
transformer	W4A4	17.03	97.45%	300	50
transformer_ref	W4A4	17.03	97.45%	300	50
text_encoder	W4A4	18.55	95.80%	448	0
vae	source FP32 copy	9.70	exact source copy	0	0
audio_vae	source FP32 copy	0.56	exact source copy	0	0

Input/output projections, time/context/refiner boundaries, embeddings, norms, and language-head boundaries excluded by the pinned H3/Qwen policy remain in source precision. “Four bit” describes eligible packed linear weights, not every tensor in the architecture.

Validation and provenance

• OrbitQuant 0.9.2 revision cd58b4ecf77f22b8c4116b3d0b7d4af258e16ba3.
• Diffusers revision abc5e9bf71fd38f53cd471bc3acaa84bc5ecbfdc.
• The live ComfyUI workflow reached terminal pass through /prompt and produced the standard VIDEO output.
• All 23 denoiser checkpoints and the final latent were persisted before source-FP32 decode.
• Visual/audio VAE weight SHA256 values match the pinned source revision; see validation/source_component_copy_audit.json.
• Exact artifact hashes are in SHA256SUMS.
• Full machine-readable profile and media evidence is in comfyui/report.json.

License and modifications

The original MiniMax H3 Community License Agreement is copied as LICENSE. See NOTICE, MODIFICATIONS.md, and the upstream QA-about-License.