Stanford's Hazy Research unveils NCM hierarchical text generation

Hazy Research at Stanford introduced Next-Concept Modeling, which builds a hierarchy of discrete codes from a coarse representation to fine details instead of generating tokens sequentially. The team reports 10–31× faster inference than evaluated diffusion baselines and roughly 4–5× fewer training tokens. The code is released under the MIT license.

• Inference is 10–31× faster than evaluated diffusion baselines
• Training uses roughly 4–5× fewer tokens
• Architecture pairs a multiscale Residual VQ-VAE tokenizer with a Transformer
• Code released in October 2026 under the MIT license