MiniMax H3 Ref2VA (mxfp8 & fp16attn_fp8)

Quantized versions of the MiniMax H3 Ref2VA (reference-to-video-audio)

diffusion model. Both keep the text-conditioning path (token_refiner,

conditioning projections, audio heads) in full precision.

Latest comfy version improves audio quality at low steps.

_minimal 5 step (talking heads, less motions) | balanced 8-steps (shift 8) | Action-rich scenes above 12 (shift max 12)_

Your browser does not support the video tag.

| File | Download | Note |

|---|---|---|

| minimax_h3_ref2va_fp16attn_fp8.safetensors | Download 28.7GB | FP16 attention + FP8-E4M3 MLP |

| minimax_h3_ref2va_mxfp8.safetensors | Download 47.6GB | E4M3 weights, E8M0 block scales |

Comparison

| | fp16attn_fp8 | mxfp8 |

|---|---|---|

| Size | 28.7 GB | 47.6 GB |

| Base checkpoint | pruned fp8_scaled (curve-form, adaln_t_table) | full bf16 (time-embedder) |

| Attention (qkv/out) | FP16 (exact) | MXFP8 (block-quantized) |

| MLP (fc1/fc2) | FP8-E4M3 per-tensor absmax | MXFP8 block-quantized |

| Quant scale granularity | per-tensor | per-32 block (better outlier handling) |

| input_scale | calibrated (fc1), fpmm on fc2 | 1.0 identity (150 layers) |

| token_refiner / conditioning | full precision | full precision |

| Relative precision | higher than pruned fp8, attn lossless | highest of the two (finer scales) |

| Memory footprint | lighter (faster load, less streaming) | heavier |

| Best for | everyday use / speed-first setups | maximum quality per weight bit |