Hugging Face's logo

Comfy-Org
/
MiniMax-H3

Diffusion Single File
comfyui
Model card Files
xet
 Community
74

[Release] MiniMax-H3-Rung-Omega (14.96 GiB, 4-Trit Base-243) + TheoremFlow-Omega (8.92x Speedup: 625s -> 70s)

#73
by yeowon - opened
Discussion
yeowon
•
edited 3 days ago

We have released yeowon/MiniMax-H3-Rung-Omega — a 50-Block Pure Rung Ω (4-Trit Balanced Base-3, Base-243) quantized checkpoint (14.96 GiB Joint-tANS / 16.22 GiB Raw) + TheoremFlow-Ω (8.92x end-to-end inference acceleration engine) for MiniMax-H3 FL2VA in ComfyUI!

Benchmark Overview

Pareto Frontier

Key Highlights

  • Smaller & More Accurate than Official #1 Recommended int8_convrot (19.53 GiB): Saves 4.57 GiB (-23.4%) (19.53 GiB → 14.96 GiB Joint-tANS / 16.22 GiB Raw uint8) while achieving +5.20%p higher Video Spatial Top-1 (78.12% vs 72.92%) and +4.17%p higher Audio Patch Top-1 (81.77% vs 77.60%) against the BF16 Teacher (37.46 GiB).
  • TheoremFlow-Ω (8.92x Speedup): Accelerates 20-step Video (864×480, 124 frames @ 24fps) + 32kHz Stereo Audio generation from 625.4s → 70.11s while preserving 99.58% Video Latent Cosine Similarity (20.80 dB SNR).
  • Drop-in ComfyUI Support: Includes src/register_comfyui_rung_omega.py, src/theoremflow_omega.py, and src/generate_minimax_h3_rung_omega_video.py.
  • Multilingual Documentation: Available in 🇺🇸 English, 🇰🇷 한국어, 🇨🇳 中文, and 🇯🇵 日本語.

🔗 Model & Demo Videos: https://huggingface.co/yeowon/MiniMax-H3-Rung-Omega
☕ Support & Priority Model Quantization Requests: https://buymeacoffee.com/tripleyoung

Sign up or log in to comment