realrebelai/FastH3-V2_GGUFs

FastH3 V2 GGUFs

GGUF quantizations of FastVideo FastH3 8-Step V2 for use with ComfyUI / ComfyUI-GGUF.

These GGUFs were converted from the full official BF16 FastH3 V2 transformer weights released by FastVideo.

These are NOT quantizations of the pruned ComfyUI checkpoint.

Model

Original model:
https://huggingface.co/FastVideo/FastVideo-FastH3-8-Step-V2

FastH3 8-Step V2 is a distilled MiniMax-H3 checkpoint from FastVideo designed to generate synchronized video and audio using only 8 transformer forwards.

The official V2 release uses:

• 35B parameter MiniMax-H3 architecture
• 8-step / 8-forward generation
• Data-Free DMD2 distillation
• VSA-H3 sparse attention
• 80% attention sparsity
• Synchronized video + audio generation

Available GGUFs

Quant	Description
Q3_K_M	Lowest VRAM/RAM option. Maximum compression.
Q4_K_M	Recommended balance between size and quality.
Q5_K_M	Higher-quality quant with increased memory requirements.

Additional quantizations may be added later.

Full-Weight Conversion

The official FastH3 V2 transformer is distributed as a large sharded BF16 checkpoint.

For these releases, the complete transformer was reconstructed and converted to the tensor layout expected by ComfyUI before GGUF quantization.

The conversion preserves the FastH3 architecture, including:

• All 50 main transformer blocks
• Token refiner blocks
• Fused QKV projections
• VSA to_gate_compress weights
• SwiGLU MLP weights
• AdaLN projections
• Video projection layers
• Audio projection layers
• Conditioning projections
• Time embeddings
• RoPE data

High-precision projection/conditioning tensors are preserved as F16/F32 where appropriate rather than indiscriminately quantizing every tensor.

ComfyUI Installation

ComfyUI-GGUF is required.

Place the downloaded GGUF in:

ComfyUI/
└── models/
    └── unet/
        └── FastH3-8-Step-V2-Q4_K_M.gguf

Restart ComfyUI or refresh your model list after installing the model.

Load the model using your GGUF unet loader.

Recommended Generation Settings

FastH3 V2 is an 8-step model.

Start with:

Steps:     8
Sampler:   Res_Multistep
Scheduler: Simple
Denoise:   1.0

The official FastVideo V2 checkpoint uses a trained video scheduler shift of 10.

Exact workflow settings may depend on the current ComfyUI MiniMax-H3 implementation and any attention/VSA patches being used.

Required Supporting Models

The FastH3 V2 GGUF contains the transformer only.

You will still need the supporting MiniMax-H3 models required by the ComfyUI FastH3 workflow:

• Qwen3-VL text encoder
• MiniMax-H3 Video VAE
• MiniMax-H3 Audio VAE for audio generation

⚠️ GGUF Text Encoder / WAN Architecture Users

If you are running the workflow through the WAN architecture implementation and want to use a GGUF Qwen3-VL text encoder, use the MiniMax-H3 GGUF encoders from my MiniMax-H3 repository:

https://huggingface.co/realrebelai/MiniMax-H3_GGUFs/tree/main

Available GGUF encoders include:

• qwen3vl-32B-MiniMax-H3-Q2_K.gguf — lower memory usage
• qwen3vl-32B-MiniMax-H3-Q4_K_M.gguf — recommended for better quality if you have enough RAM/VRAM

Place the encoder in:

ComfyUI/
└── models/
    └── text_encoders/
        └── qwen3vl-32B-MiniMax-H3-Q4_K_M.gguf

These are the MiniMax-H3-specific Qwen3-VL GGUF encoders intended for this architecture. If you are using the WAN architecture path with GGUF encoder loading, use one of these rather than a generic Qwen3-VL GGUF.

VAE Files

You will also need the MiniMax-H3 Video VAE and Audio VAE.

Official ComfyUI-compatible supporting files can be found here:

https://huggingface.co/Comfy-Org/MiniMax-H3/tree/main/vae

Both VAEs are required if you want synchronized video + audio generation.

Q4_K_M

Q4_K_M is the recommended starting point for most users.

It significantly reduces the storage and memory requirements of the original ~35B BF16 transformer while retaining substantially more of the original model than lower-bit quantization.

Users with more available system/GPU memory may prefer Q5_K_M, while Q3_K_M is intended for more memory-constrained systems.

Important: VSA / Attention

FastH3 V2 was trained using VSA-H3 sparse attention at 80% sparsity.

Attention implementations, VSA patches, alternative attention backends, and post-processing/upscaling nodes can affect output quality and temporal consistency.

When troubleshooting generation quality, test the raw FastH3 output before adding video restoration, interpolation, or upscaling stages.

Scope

The official FastH3 V2 checkpoint supports text-to-audio-video generation.

According to FastVideo, FL2VA and Ref2VA were not distilled for this checkpoint.

Quality on difficult motion, very fine details, and some audio scenarios may remain below the original MiniMax-H3 base model.

Credits

All model architecture, training, distillation, and original weights belong to their respective authors.

FastH3 / FastVideo:
https://github.com/hao-ai-lab/FastVideo

Original FastH3 V2 model:
https://huggingface.co/FastVideo/FastVideo-FastH3-8-Step-V2

Original base model:
https://huggingface.co/MiniMaxAI/MiniMax-H3

GGUF conversions:
Rebel AI / realrebelai

License

These quantizations inherit the licensing requirements and usage restrictions of the original FastH3 V2 / MiniMax-H3 model.

Please review the original model license before using or redistributing these files:

https://huggingface.co/FastVideo/FastVideo-FastH3-8-Step-V2

---

If these GGUFs helped you run FastH3 V2 on hardware that couldn't comfortably handle the original BF16 transformer, consider dropping a ❤️ on the repository.