ModelScope releases EntroPack, a PyTorch library for fractional-bit weight compression

AIModelScope has open-sourced EntroPack, a PyTorch tensor compression library that quantizes models at arbitrary bits per parameter rather than fixed 4-bit or 8-bit precisions. In one example on a 16G GPU, Qwen-Image-2.1 quantized at 4.8 bpp uses 15.86 GiB and cuts reconstruction error from NF4's 9.40% to 4.26%. Pre-quantized weights for Z-Image-Turbo, Qwen-Image-2.1, and MiniMax-H3 at 4 to 8 bpp are available for DiffSynth-Studio and ComfyUI, and the library is licensed under Apache 2.0.

AIModelScope announces EntroPack, a PyTorch tensor compression library that lets users choose fractional bits per parameter instead of fixed 4-bit or 8-bit formats. In a Qwen-Image-2.1 test at 1024 x 1024, targeting 4.8 bpp peaked at 15.86 GiB and cut weight reconstruction error from NF4's 9.40% to 4.26%. Pre-quantized weights for Z-Image-Turbo, Qwen-Image-2.1, and MiniMax-H3 are available, and the library code is under Apache 2.0.