NCM weights land with a 31x text-generation speed claim
Hazy Research's new checkpoint release makes coarse-to-fine text generation available to inspect, with benchmark-specific speed claims and an incomplete OpenWebText package.
A different route to parallel text generation
Hazy Research has made Next-Concept Modeling checkpoints public, giving researchers weights and code for a text generator that builds a sequence from a coarse sketch to finer detail. The Hugging Face history records a public-release commit at 05:29:28 UTC on October 14. The repository was created October 13; the underlying paper was submitted to arXiv on October 9.
NCM first uses a learned tokenizer to turn text into a hierarchy of discrete codes. A transformer then predicts one level at a time, generating all positions within that level in parallel. The tokenizer’s decoder converts the completed representation back into text. This differs from producing one token per step or repeatedly denoising the whole sequence.
What the 31x figure actually compares
The authors report that NCM matches the strongest diffusion baseline’s ROUGE-L on XSum summarization while generating 31 times faster. They also report a 21-fold speedup over DUO on LM1B with better generative perplexity. These are the team’s experimental results, not an independently reproduced performance verdict.
The paper generally evaluates generators around 100 million parameters. Its wall-clock protocol generates 1,024 sequences at batch size 16 on one NVIDIA H100 SXM 80GB GPU. Those conditions matter: the figures do not establish a universal speedup over production chat models, every diffusion system or a consumer GPU setup.
Weights to inspect, with one visible gap
The release includes tokenizer and generator checkpoints for XSum, WMT14 German-English translation, Sudoku, LM1B and TinyStories, alongside configurations and evaluation instructions. The code repository declares an MIT license and recommends Python 3.10 or newer with CUDA-capable PyTorch.
There is an availability caveat. Although the project README lists OpenWebText in its model table, the inspected checkpoint revision contains only owt/TODO.md for that task, without its tokenizer or generator weights. The immediate value is access to the released task-specific experiments and their evaluation path. This is research material for examining an alternative generation architecture, rather than a ready-made replacement for a general-purpose assistant.