MiniMax H3 Ref2VA — pruned scaled FP8
ComfyUI mixed-precision FP8 export of our pruned BF16 checkpoint. Weight storage is 20.96 GB, down from 40.23 GB. The existing AdaLN pruning is preserved; no further structural pruning is applied.
Quantization policy
| Layers, across 50 transformer blocks | Stored weights | Matrix multiplication |
|---|---|---|
| Attention QKV and output, MLP fc1 (150 matrices) | Scaled FP8 E4M3FN | FP8 weights and statically scaled FP8 activations |
| MLP fc2 (50 matrices) | Scaled FP8 E4M3FN | Weights dequantized; BF16 GEMM in our run |
Each matrix has one FP32 weight scale, computed as max(abs(weight)) / 448. Conversion divides by that scale and rounds to E4M3FN. Per-layer comfy_quant tags and the 150 published static activation scales are reused from Comfy-Org. We have not independently calibrated activations and do not know the upstream calibration dataset. This reproduces the observable format and scaling policy, not an unpublished upstream conversion procedure or a guarantee of bit-identical FP8 weights.
All 332 other tensors are preserved byte for byte from our pruned source, including BF16 AdaLN projections and the FP32 timestep table. Norms, token refiners and input/output projections are excluded from FP8. ComfyUI may cast nonquantized parameters to its configured compute dtype when loading; storage preservation does not imply all computations retain their storage dtype. Attention uses BF16 PyTorch SDPA. Text encoder and VAEs are separate and are not quantized by this export.
Sources
- Original MiniMax H3:
MiniMaxAI/MiniMax-H3at42ed227ee7df40d41602854ae760620d6eb651fe. - Pruned BF16 source:
haideraqeeb/MiniMax-H3-Ref2VA-Pruned-BF16at4fb0e01948845d44d0a901810f3e2400c82841b4. - FP8 configuration/scales:
Comfy-Org/MiniMax-H3ate5eb578a89295337b8ff433a035929ce0279e0b6,minimax_h3_ref2va_pruned_fp8_scaled.safetensors. - Validated ComfyUI:
e377e263049f9338b4d12a3dd417b36ae62948ff, PyTorch 2.11.0+cu130, RTX PRO 6000 Blackwell.
Loading
Put the checkpoint in ComfyUI/models/diffusion_models/ and select it in Load Diffusion Model using an H3-compatible ComfyUI with mixed-precision comfy_quant support. Use the H3 Ref2VA conditioning nodes and separate text encoder/video/audio VAEs. Keep automatic/default weight handling to use the embedded quantization configuration. Native FP8 execution was verified on the hardware above; other hardware may use different paths.
Validation and five-second comparison
Compared with the pruned BF16 source using the same official Ref2VA example: 1344 × 768, 30 steps, res_multistep, beta schedule, seed 0, no LoRA, identical cached conditioning, BF16 text encoder and FP32 video/audio VAEs. Initial noise and sigma schedules match exactly. Native output is 124 frames at 24 fps; presentation clips retain the first 120 frames and matching stereo audio.
All 200 weight checks passed and all 332 excluded tensors match the source. The runtime audit observed 4500 successful native FP8 matrix multiplications with zero kernel errors. Across 150 activation-quantized layers, 21,423 of 2,233,324,800,000 audited values exceeded the saved representable ranges (fraction 9.5924247e-09; 39 layers affected). See the per-layer audit for distribution.
| Difference from pruned BF16 | Value |
|---|---|
| Decoded video pixel PSNR | 22.75 dB |
| Audio waveform relative L2 | 0.48404 |
| Video latent relative L2 | 0.33587 |
| Audio latent relative L2 | 0.27450 |
These are divergence measurements, not perceptual quality scores. One prompt and seed do not establish general quality equivalence. FP8 timing and memory measurements include activation-audit overhead and are not a clean speed benchmark. Supporting JSON reports are in validation/. Scripts record construction but currently use the documented local workspace paths; they are not a portable CLI.
AI-generated examples:
License
Powered by MiniMax H3. The MiniMax H3 Community License Agreement applies; see LICENSE and NOTICE. This is an independently modified checkpoint, not an official MiniMax release.