brurpo/MiniMax-H3-NVFP4-GGUF

MiniMax H3 NVFP4 GGUF

Native GGML_TYPE_NVFP4 GGUF diffusion-model weights for MiniMax H3. These files target the MiniMax H3 support in stable-diffusion.cpp and Blackwell-class NVIDIA GPUs.

This repository contains all four combinations of task and pruning level:

File	Task	Bytes	NVFP4 tensors	SHA-256
minimax_h3_fl2va_nvfp4.gguf	FL2VA, full	18,685,049,568	470	40699037ee725b5fab15594694bbbd50abbaec65bf3c03420be3be171105e229
minimax_h3_fl2va_pruned_nvfp4.gguf	FL2VA, pruned	11,380,270,144	418	7400c9a10b667322a13cf9f874182e9f5f0daf9b7129eeeef3d90b65b56e176e
minimax_h3_ref2va_nvfp4.gguf	Ref2VA, full	18,685,049,568	470	4ea6459d6282b3be3a60274ba623431869d6d2299a470ad1fc8888997066e7dd
minimax_h3_ref2va_pruned_nvfp4.gguf	Ref2VA, pruned	11,380,270,144	418	aa3d1b849da4014eb2f09a07c09c99a416bb16499493bd22778e13e92fba32bb

Compatibility

These are GGUF files with native NVFP4 tensors. They are not ComfyUI packed-NVFP4 safetensors. They were converted and tested with stable-diffusion.cpp commit ea7f0c8.

The diffusion model is not a complete standalone H3 package. In stable-diffusion.cpp it also needs:

• a MiniMax H3 Qwen3-VL text encoder, such as qwen3vl_32b_minimax_h3-Q2_K_M.gguf;
• minimax_h3_video_vae_fp16.safetensors;
• minimax_h3_audio_vae_fp32.safetensors when generating synchronized audio.

NVFP4 GPU execution requires a compatible Blackwell NVIDIA GPU and a CUDA build whose compiled architectures include the installed device. Other backends may be able to load the GGUF container but are not the target of this release.

Provenance

The source tensors are the official BF16 releases under Comfy-Org/MiniMax-H3:

Source file	Bytes	Official LFS SHA-256
diffusion_models/minimax_h3_fl2va_bf16.safetensors	66,280,487,368	907d4add438438ec1544f5240c3b38532ed934fe6be75677a6bbda2a6fdd6182
diffusion_models/minimax_h3_fl2va_pruned_bf16.safetensors	40,225,724,176	a32572fb90b5508b201ec7c2eddcc184b13ddfd3c6f6d2cf06a0b46535d541b4
diffusion_models/minimax_h3_ref2va_bf16.safetensors	66,280,487,368	e32c54c1a7b4f5f397f195cea267ccb18806303bb665678c4bee60953bdf3026
diffusion_models/minimax_h3_ref2va_pruned_bf16.safetensors	40,225,724,176	37c0da793e20ca735272ec2be655f08a2e10f97a3ec8fdfb40f5b39a736ed6fe

Conversion command pattern:

sd-cli --mode convert --model SOURCE.safetensors --output OUTPUT.gguf --type nvfp4

Validation

All four files passed the following checks before upload:

1. official source byte length and LFS SHA-256;
2. GGUF v3 structure, expected H3 tensor count, and native type-40 NVFP4 tensor inventory;
3. complete stable-diffusion.cpp load with the Q2 H3 text encoder plus video and audio VAEs;
4. one-step, 256x256, 9-frame CUDA video generation on an NVIDIA GeForce RTX 5070 Ti (compute capability 12.0);
5. successful diffusion sampling, VAE decode, and non-empty AVI output.

The full variants exceed 16 GiB if staged as one buffer. They passed with --max-vram 14 --stream-layers; the pruned variants passed with ordinary CPU parameter offload. Low-level backend tests also identified CUDA0 and passed native NVFP4 matrix multiplication plus the supported SM 12.0 flash-attention cases.

License

The upstream MiniMax H3 Community License Agreement is included as LICENSE. Review it before downloading, redistributing, or using these derivatives.

Related artifact

An independently published compatible NVFP4 GGUF exists at convertor/minimax-h3-gguf for the Ref2VA-pruned variant. This repository's purpose is to provide a consistently converted and validated four-variant set.