berryber09/MiniMax-H3-FastH3-VSA-4step-pruned-w4a8

MiniMax-H3 FastH3 VSA 4-step (fl2va, pruned) — W4A8

W4A8 quantization of FastH3 4-step Preview v1 VSA (DataFree) (FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree, checkpoint step 1300), converted to the ComfyUI fl2va pruned layout and quantized from the bf16 conversion.

• Format: asym_w4a8_int8, group_size=16, per-tensor Lloyd-Max codebook + fp8 group scales (Kijai / comfy-kitchen W4A8 layout).
• Size: ~16.4 GB.
• Layers: 200 2D-linear layers quantized. Kept in bf16: the 50 VSA gate layers (to_gate_compress, 3.85 GB; the quantizer policy predates VSA and does not cover them), the token refiner, adaLN projections and first/last layers (382 tensors kept).
• Quantizer: comfyui-mixed-quantizer --format w4a8 --group-size 16 --codebook-mode fit; validation 65 passed, 0 failed; max sampled relL2 0.073.
• Verified: loads via UNETLoader and generates coherent video in ComfyUI at 4 steps with VSA attention.
• sha256: 9fca42ce867187e2e0ba7e7ad451e5dfe03cb5c0fee10926c730b3c4338ce05b

Usage

4 steps, cfg 1. VSA (video sparse attention) is enabled by ComfyUI on models with to_gate_compress weights.

Requirements

• ComfyUI ≥ v0.31.0 (native W4A8 loader).
• comfy-kitchen with AsymW4A8Int8Layout.
• CUDA SM ≥ 8.0.

License

Inherits the MiniMax H3 Community License of the base model, including its territory restrictions. Read it before use.

Same-seed A/B (1280x720, 5 s, 4 prompts, RTX 3090)

Against our regular H3 fl2va w4a8 at its turbo settings with block-sparse attention: sampler 57.5 s vs 80.1 s per clip (VAE decode equal, 36.8 s). Quality on par: more dynamic motion and crisper detail here; the regular build had better light on the landscape and more natural skin in the dialogue close-up. Audio present on all clips; dialogue transcribed word for word on both.