Stanford and Together AI's NCM generates text 10–31× faster than diffusion models

Stanford's Hazy Research lab and Together AI introduced Next-Concept Modeling, which builds text hierarchically from outline to detail and predicts each level in parallel instead of one token at a time. Tests show 10–31× faster generation than strong diffusion language models with 4–5× fewer training tokens, though the largest model tested has only 725M parameters.

• Generation is 10–31× faster than diffusion language models in the authors' tests
• Training needs 4–5× fewer tokens, including tokenizer training
• Largest model tested is 725M parameters; no hosted API, code on GitHub
• Sequential steps depend on the number of levels, not output length