---

license: other license_name: minimax-h3-community license_link: https://huggingface.co/MiniMaxAI/MiniMax-H3/raw/main/LICENSE tags:

• text-to-video
• image-to-video
• gguf
• comfyui
• minimax-h3
• singularity
• quantized base_model: WarmBloodAban/Minimax-h3_Singularity

---

Minimax-H3 Singularity V2 (Pruned) - GGUF Quantizations

Quantized GGUF versions of WarmBloodAban/Minimax-H3 Singularity, optimized for ComfyUI-GGUF.

---

📁 Available GGUF Files & Sizes

Filename	File Size	Quantization Type	Recommended VRAM / Use Case
Minimax-h3-Singularity-ref2va-v2-Pruned-Q3_K_M.gguf	8.9 GB	Q3_K_M (3-bit Medium)	Low-VRAM systems (12 GB–16 GB VRAM)
Minimax-h3-Singularity-ref2va-v2-Pruned-Q4_K_M.gguf	11.6 GB	Q4_K_M (4-bit Medium)	Sweet Spot for standard GPUs (16 GB–24 GB VRAM)
Minimax-h3-Singularity-ref2va-v2-Pruned-Q5_K_M.gguf	14.1 GB	Q5_K_M (5-bit Medium)	High fidelity for 24 GB+ VRAM
Minimax-h3-Singularity-ref2va-v2-Pruned-Q6_K.gguf	16.7 GB	Q6_K (6-bit)	Near-lossless quality

---

🛠️ Engineering & Pruning Details

• Curve-Pruned AdaLN Architecture: The 51 unpruned AdaLN modulation modules ([Out, 2688]) from the original 34GB checkpoint were projected onto the canonical rank-8 Ref2VA subspace (adaln_basis + adaln_mean), reducing parameters by 13.03 Billion and saving ~26 GB of disk space.
• Lossless Precision Guards: All normalization layers (RMSNorm, LayerNorm), timestep embeddings (adaln_t_table), and modality input/output heads (video_patch_proj, audio_patch_proj, final_layer.video_out, final_layer.audio_out) are preserved in lossless FP32.
• ConvRot Un-Rotation: Source INT8 weights were accurately un-rotated using ComfyUI's exact Regular Hadamard ($H_4^{\otimes 4}$) transform in float64 before K-quantization, eliminating quantization artifacts and color shifts.
• LoRA Compatibility: Includes the canonical adaln_affine.safetensors curve basis, ensuring 100% compatibility with the ComfyUI pruned H3 LoRA and PowerLoRA stack ecosystem.

---

🚀 How to Use in ComfyUI

1. Place your chosen .gguf file into your ComfyUI models folder: ComfyUI/models/diffusion_models/
2. Load the model using UnetLoaderGGUF.
3. Pair it with the Qwen3-VL text encoder and MiniMax-H3 Video + Audio VAEs.

---

📜 Acknowledgments & License

• Original Model: WarmBloodAban/Minimax-H3_Singularity
• Base Architecture: MiniMax H3 by MiniMax, licensed under the MiniMax H3 Community License Agreement.
• GGUF Conversion & Pruning: By Abiray.