HOT RADAR
opensourceHEAT 8.2°

AI CURATED HOT STORY · 11/19/2026

ModelScope Open-Sources EntroPack for Arbitrary-Rate Weight Compression

1 reports archived1 independent sourcesupdated 11/19/2026, 19:15:00
Synthesis & Latest Updates
1 Sources Cross-Validated

ModelScope has open-sourced EntroPack, a weight compression scheme that folds information entropy into joint quantization and compression design, turning bitrate control into entropy regulation of quantized outputs and enabling arbitrary bitrates and multiple numeric precisions. It requires no activation calibration or fine-tuning, runs on NVIDIA, AMD and ZW-PPU hardware, fits image models into 10GB of VRAM, and cuts quantization error from 9.4% to 4.26%. The implementation is fully open and integrated with diffusion and LLM ecosystems.

LATEST/ModelScope has open-sourced EntroPack, a weight compression scheme that folds information entropy into joint quantization and compression design, turning bitrate control into entropy regulation of quantized outputs and enabling arbitrary bitrates and multiple numeric precisions. It requires no activation calibration or fine-tuning, runs on NVIDIA, AMD and ZW-PPU hardware, fits image models into 10GB of VRAM, and cuts quantization error from 9.4% to 4.26%. The implementation is fully open and integrated with diffusion and LLM ecosystems.

Hourly heat curve

17 fully observed hours
Now
8.2
Peak
11.820:00
24h change
–

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

Coverage timeline

Total 1 reports · Latest first
  1. 量子位 (微信公众号)T2·74 pts
    • Bitrate control is reframed as entropy regulation of quantized outputs, supporting arbitrary bitrates and multiple precisions
    • No activation calibration or fine-tuning needed; runs on NVIDIA, AMD and ZW-PPU GPUs
    • Image models fit in 10GB VRAM with quantization error cut from 9.4% to 4.26%; integrated with diffusion and LLM ecosystems