model card: single-pass build, retract double-quant caveat

minimax_h3_fl2va_pruned_nvfp4.safetensors: rebuild from pruned_bf16 (single-pass NVFP4)

minimax_h3_ref2va_pruned_nvfp4.safetensors: rebuild from pruned_bf16 (single-pass NVFP4)

add rockerBOO mixed-precision convrot_int8 files, with credit

mirror minimax_h3_fl2va_pruned_nvfp4_convrot_int8.safetensors from rockerBOO/minimax-h3-nvfp4 (credit: rockerBOO)

mirror minimax_h3_ref2va_pruned_nvfp4_convrot_int8.safetensors from rockerBOO/minimax-h3-nvfp4 (credit: rockerBOO)

Soften NVFP4 recommendation; scope measured table to speed/size only

Honest limitations: int8_convrot may beat NVFP4 on motion quality; retract prior no-degradation claim; add VRAM ceiling

Document fl2va_pruned_nvfp4

Add minimax_h3_fl2va_pruned_nvfp4.safetensors (requantized from diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors)

Upload README.md with huggingface_hub

Upload pruned_to_nvfp4.py with huggingface_hub

Upload minimax_h3_ref2va_pruned_nvfp4.safetensors with huggingface_hub

Add prompting guide: H3 expects a structured IR, not prose

Upload vae/minimax_h3_audio_vae_fp32.safetensors with huggingface_hub

Upload vae/minimax_h3_video_vae_fp16.safetensors with huggingface_hub

Upload text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors with huggingface_hub

Upload README.md with huggingface_hub

Upload minimax_h3_ref2va_nvfp4_full.safetensors with huggingface_hub

Upload minimax_h3_ref2va_nvfp4_mixed.safetensors with huggingface_hub

Upload make_profile.py with huggingface_hub

Upload profiles/minimax_h3_nvfp4_full.json with huggingface_hub

Upload profiles/minimax_h3_nvfp4_mixed.json with huggingface_hub

Upload README.md with huggingface_hub

initial commit