Next-Concept Modeling: 9–31× faster than diffusion baselines

HazyResearch released Next-Concept Modeling (NCM), a two-stage parallel text generation framework that beats or matches strong diffusion baselines on different tasks and datasets while generating 9–31× faster. NCM pairs a multiscale residual VQ tokenizer, which represents a sequence as a hierarchy of discrete codes, with a transformer that predicts that hierarchy coarse to fine, generating positions within a level in parallel for one forward evaluation per level. Weights and tokenized datasets for XSum, WMT14 De–En, Sudoku (easy, medium, hard), LM1B (small, medium, large) and TinyStories are hosted in the NCM Hugging Face collection, with each released model consisting of a frozen tokenizer and a next-concept generator.

Fast, Parallel Coarse-to-Fine Text Generation Next-Concept Modeling NCM is a two-stage framework for parallel text generation. A multiscale residual VQ tokenizer represents a sequence as a hierarchy of discrete codes, and a transformer predicts that hierarchy from coarse to fine. Positions within a level are generated in parallel, so generation requires one forward evaluation per level. NCM generates coarse to fine: each level predicts a longer sequence in parallel, and decoding every level shows a sample being refined from a rough sketch to the full output in a few steps. NCM beats or matches strong diffusion baselines on different tasks & datasets while generating 9-31× faster. Weights and tokenized datasets are hosted in the NCM Hugging Face collection https://huggingface.co/collections/hazyresearch/ncm . Each released model consists of a frozen tokenizer and a next-concept generator. | Task | Model | Tokenizer | Generator | Config | Data | |---|---|---|---|---|---| | XSum | – | tokenizer.pt https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/xsum/tokenizer.pt | generator.pt https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/xsum/generator.pt | config.yaml https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/xsum/config.yaml | test set https://huggingface.co/datasets/hazyresearch/ncm-datasets/tree/main/xsum | | WMT14 De–En | – | tokenizer.pt https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/wmt14/tokenizer.pt | generator.pt https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/wmt14/generator.pt | config.yaml https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/wmt14/config.yaml | test set https://huggingface.co/datasets/hazyresearch/ncm-datasets/tree/main/wmt14 | | Sudoku | easy | tokenizer.pt https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/sudoku/easy/tokenizer.pt | generator.pt https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/sudoku/easy/generator.pt | config.yaml https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/sudoku/easy/config.yaml | test boards https://huggingface.co/datasets/hazyresearch/ncm-datasets/tree/main/sudoku | | | medium | tokenizer.pt https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/sudoku/medium/tokenizer.pt | generator.pt https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/sudoku/medium/generator.pt | config.yaml https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/sudoku/medium/config.yaml | | | | hard | tokenizer.pt https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/sudoku/hard/tokenizer.pt | generator.pt https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/sudoku/hard/generator.pt | config.yaml https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/sudoku/hard/config.yaml | | | LM1B | small | tokenizer.pt https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/lm1b/tokenizer.pt | generator.pt https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/lm1b/small/generator.pt | config.yaml https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/lm1b/small/config.yaml | tokenized text https://huggingface.co/datasets/hazyresearch/ncm-datasets/tree/main/lm1b | | | medium | | generator.pt https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/lm1b/medium/generator.pt | config.yaml https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/lm1b/medium/config.yaml | | | | large | | generator.pt https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/lm1b/large/generator.pt | config.yaml https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/lm1b/large/config.yaml | | | TinyStories | – | tokenizer.pt https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/tinystories/tokenizer.pt | generator.pt https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/tinystories/generator.pt | config.yaml https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/tinystories/config.yaml | tokenized text https://huggingface.co/datasets/hazyresearch/ncm-datasets/tree/main/tinystories | | OpenWebText | – | tokenizer.pt https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/owt/tokenizer.pt | generator.pt https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/owt/generator.pt | config.yaml https://huggingface.co/hazyresearch/ncm-checkpoints/blob/main/owt/config.yaml | tokenized text https://huggingface.co/datasets/hazyresearch/ncm-datasets/tree/main/owt | git clone https://github.com/HazyResearch/ncm.git cd ncm pip install -r requirements.txt Python 3.10 or newer and a CUDA-capable PyTorch installation are recommended. Download the checkpoints and evaluation inputs: hf download hazyresearch/ncm-checkpoints --local-dir artifacts hf download hazyresearch/ncm-datasets --repo-type dataset --local-dir data Then run: python evaluate.py --config config/eval/<task .yaml --checkpoint-dir artifacts/<task --data-dir data/<task <task is one of xsum, wmt14, sudoku-easy, sudoku-medium or sudoku-hard python evaluate.py --config config/eval/<dataset .yaml --checkpoint-dir artifacts/<dataset <dataset is tinystories, owt or lm1b-<size with <size one of small, medium or large This runs one generation shard per visible GPU, merges them, and prints ROUGE-1/2/L and generative perplexity for XSum; cased and lowercased sacreBLEU, chrF and length ratio for WMT14; solved, cell and valid-grid rates for Sudoku; and generative perplexity under GPT-2 large with unigram entropy, over 5 seeds of 256 unconditional samples, for TinyStories, OpenWebText and LM1B, the latter under the FLM scoring protocol. Add --cfg <scale for classifier-free guidance on XSum and WMT14; the guided WMT14 row uses --cfg 1.0 . - evaluate.py : evaluation entry point - scripts/ : per-GPU generation shard and the merge that scores it - config/eval/ : one config per released checkpoint, identical to the config.yaml shipped with it - models/ , tokenization/ : generator and tokenizer implementations @misc{kumbong2026nextconcept, title={Next-Concept Modeling: Fast, Parallel Coarse-to-Fine Text Generation}, author={Hermann Kumbong and Ishan S. Khare and Iddah Mlauzi and Andrew Woen and Charles Ding and E. Kelly Buchanan and Simran Arora and Azalia Mirhoseini and Christopher Ré}, year={2026}, eprint={2610.14162}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2610.14162}, } Released under the MIT license https://github.com/HazyResearch/ncm/blob/main/LICENSE .