This New AI Architecture Is 31x Faster (NCM)

Next-Concept Modeling (NCM) from Stanford’s Hazy Research writes an answer as a rough sketch first and the exact words last, in a handful of passes instead of one pass per token. I ran the released models on a DGX Spark and timed NCM against a same-size word-by-word LLM: it’s up to 4x faster for one request and 4–8x slower at 64 requests. I also trained my own diffusion model, which tied NCM’s 86% on hard Sudoku once it filled the most confident cell first. Paper: https://arxiv.org/abs/2610.14162 Code: https://github.com/HazyResearch/ncm Checkpoints: https://huggingface.co/hazyresearch/ncm-checkpoints Hazy Research blog: https://hazyresearch.stanford.edu/blog/2026-10-12-ncm Sudoku setup (S-FLM): https://github.com/jdeschena/s-flm Most-confident-first decoding (Kim et al. 2025): https://arxiv.org/abs/2502.06768 Diffusion Gemma Locally: https://youtu.be/I2-_3ieJelE Diffusion LLMs Are Here!: https://youtu.be/0B9EMddwlOQ Optional disclosure: Diffusion Gemma and the Gemini logo appear briefly. If you want the line “I’m joining Google; this is based on public information and my views are my own”, confirm the wording with your recruiter. Chapters (timed from the final edit) 0:00 What NCM (Next-Concept Modeling) is 0:55 Two questions the paper never tested 1:57 Why LLMs write one token at a time 3:07 Diffusion models: the obvious fix 4:24 How NCM works: coarse-to-fine levels 5:35 What a “concept” is 6:04 The generator: one level per pass 7:06 Paper results: 31x faster, 86% on Sudoku 7:49 What most coverage missed 8:58 Test 1: watching NCM solve Sudoku 10:01 Test 2: NCM vs a same-size LLM 10:38 64 requests: the result flips 12:11 Test 3: my own diffusion Sudoku model 12:40 Most-confident-first ties NCM 13:48 NCM limitations 14:52 Bottom line