This model uses DFloat11 lossless compression. It's 30% smaller than the original BFloat16 model, yet produces bit-identical outputs and runs efficiently on GPUs.

We apply Huffman coding to the exponent bits of BFloat16 model weights, which are highly compressible. We leverage hardware-aware algorithmic designs to enable highly efficient, on-the-fly weight decompression directly on the GPU. Find out more in our research paper.

A complete usage guide is available in our GitHub repository: https://github.com/LeanModels/DFloat11/tree/master/examples/wan2.1.