HazyResearch Claims NCM Generates Text 9 to 31 Times Faster Than Diffusion Baselines

HazyResearch publishes NCM, a two-stage text generation framework whose coarse-to-fine decoder predicts token hierarchies in parallel, with authors claiming 9-31x faster generation than strong diffusion baselines at matching quality. Code, checkpoints, and tokenized datasets are released under the MIT license on Hugging Face, requiring Python 3.10 or newer and a CUDA-capable PyTorch install.

• The paper appears as arXiv 2610.14162 under cs.CL, credited to Hermann Kumbong and Christopher Ré.
• Evaluation configs span XSum, WMT14 De-En, three Sudoku difficulties, TinyStories, OpenWebText and LM1B at three sizes.
• Each released model pairs a frozen multiscale residual VQ tokenizer with a separate next-concept generator checkpoint.