MiniMax-H3-Singularity GGUF

GGUF quantized versions of WarmBloodAban/Minimax-h3_Singularity, a community fine-tune of MiniMax H3 specialized in enhanced visual fidelity โ€” HDR clarity, distant-face restoration, cleaner skin, stronger action motion, and improved VFX. These quantizations reduce VRAM requirements for use with ComfyUI, enabling local inference on consumer GPUs.

Model Details

Attribute Value
Base Model WarmBloodAban/Minimax-h3_Singularity
Base Architecture MiniMax-H3 joint audio-video diffusion transformer
File Format GGUF
Quantization K-quant / Q8_0, sensitive layers preserved in F32
License MiniMax H3 Community License Agreement

Available Files

File Size Notes
Minimax-H3-Singularity-Q3_K_M.gguf 8.9 GB Smallest, fastest; some quality loss
Minimax-H3-Singularity-Q4_K_S.gguf 11.6 GB Good balance for 12 GB GPUs
Minimax-H3-Singularity-Q4_K_M.gguf 11.6 GB Recommended for 12 GB GPUs
Minimax-H3-Singularity-Q5_K_S.gguf 14.1 GB Better quality, 16 GB GPUs
Minimax-H3-Singularity-Q5_K_M.gguf 14.1 GB Recommended for 16 GB GPUs
Minimax-H3-Singularity-Q6_K.gguf 16.7 GB High quality, 24 GB GPUs
Minimax-H3-Singularity-Q8_0.gguf 21.6 GB Near-lossless, 24 GB+ GPUs

Recommended Usage

ComfyUI Setup

Place the downloaded .gguf file in your ComfyUI models/unet/ directory (or models/diffusion_models/ depending on your ComfyUI-GGUF fork). You will also need the following components:

Component Source
Text Encoder qwen3vl_32b_minimax_h3_Q4_K_M.gguf from Abiray/MiniMax-H3-GGUF
Video VAE minimax_h3_video_vae_fp16.safetensors from Comfy-Org/MiniMax-H3
Audio VAE minimax_h3_audio_vae_fp32.safetensors from Comfy-Org/MiniMax-H3

For accelerated generation, you can optionally pair this model with the minimax_h3_ref2v_turbo_4step_v0.1 LoRA, which enables 4-step fast inference.

VRAM Recommendations

GPU VRAM Recommended Quant
12 GB Q4_K_M
16 GB Q5_K_M
24 GB Q6_K or Q8_0

License

MiniMax H3 is licensed under the MiniMax H3 Community License Agreement, Copyright ยฉ 2026 MiniMax. Key terms include:

  • Applicable Territory: Worldwide, excluding the European Union, the United Kingdom, the Republic of Korea, and the United States of America.
  • Excluded Territories require a separate license from MiniMax.
  • Commercial use is permitted for organizations with annual revenue below $20 million, with prominent attribution.
  • The full license text is included in the LICENSE file in this repository.

These GGUF quantizations are community-produced derivatives and are not affiliated with or endorsed by MiniMax.

Acknowledgments

  • WarmBloodAban / AIGC-Singularity for the original Minimax-h3_Singularity fine-tune.
  • Comfy-Org for the MiniMax-H3 ComfyUI integration.
  • city96 for the ComfyUI-GGUF tooling used to produce these quantizations.
Downloads last month
4,378
GGUF
Model size
20B params
Architecture
wan
Hardware compatibility
Log In to add your hardware

3-bit

4-bit

5-bit

6-bit

8-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Abiray/MiniMax-H3-Singularity-GGUF

Quantized
(1)
this model

Space using Abiray/MiniMax-H3-Singularity-GGUF 1