CONFIRMED: This is a research release published on the Hazy Research lab blog on 12 October 2026. All the results are the authors’ own tests against other experimental models on small research datasets. They have not been checked independently, and NCM was not compared with ChatGPT- or Claude-class systems.
Researchers at Stanford University’s Hazy Research lab have introduced a new way for AI to write text, called Next-Concept Modeling (NCM) [1].
Instead of writing one word after another, NCM starts with a rough outline of what it wants to say and then fills in more detail in a handful of steps, working on many parts of the text at the same time [1].
The lab says that, in its tests, NCM “achieves competitive or better quality than strong diffusion baselines while generating text 10–31× faster and seeing 4–5× fewer training tokens” [1].
Three ways to write text
The lab sums up the difference like this: “Unlike autoregressive models that generate one token at a time, or diffusion models that repeatedly refine the entire sequence, NCM predicts increasingly detailed representations of text in parallel” [1].
In plain terms:
- Autoregressive models, the approach behind today’s best-known chatbots, write one token (a word or part of a word) at a time. Each token has to wait for the one before it.
- Diffusion models work on the whole passage at once and improve it over many rounds. That is more parallel, but the lab notes they “still require many sampling steps” and “process the full sequence at each step” [1].
- NCM builds the text from coarse to fine. Early levels are short and capture “the broad structure of the text”, while later levels “add finer details” [1].
How it works
NCM is trained in two stages [1]:
- A multiscale tokenizer. This turns text into a stack of levels of codes, from coarse to fine. The lab says it builds on VAR, a coarse-to-fine method used to generate images, with changes to make it work for text.
- A transformer that fills in each level. When generating, the model produces one level at a time, “predicting all codes within each level in parallel”. The tokenizer then turns the finished stack back into text.
Because the early levels are short, the lab says the first steps are “especially inexpensive” [1].
What the lab reports
All of these are the authors’ own measurements, and the comparisons are with diffusion language models unless stated [1]:
- Speed. Timed on “a single NVIDIA H100 SXM 80GB GPU”, NCM “runs 21× faster on LM1B and 10× faster on OpenWebText than the strongest diffusion baselines”. LM1B and OpenWebText are standard research text datasets.
- Summarising. It “matches the strongest diffusion baseline’s ROUGE-L on XSum while running 31× faster”. XSum is a test of writing one-sentence news summaries; ROUGE-L is a score for how closely a summary matches a reference.
- Sudoku. This is the only test that includes an autoregressive model. On hard puzzles (30 clues), the share of boards fully solved was: autoregressive 1.0%, MDLM 30.2%, DUO 58.4%, FMLM+ 81.3% and NCM 86.1%. (MDLM, DUO and FMLM+ are diffusion models.)
- Training data. On LM1B and OpenWebText, NCM used “approximately 4.9× and 4× fewer tokens than strong diffusion baselines”, counting the training of both stages.
The lab calls its scaling results, which look at what happens as the model gets bigger, “preliminary” [1].
Why speed matters
Every time an AI model answers a question it runs on costly chips, and that running cost is known as inference. The lab says language model inference “has become a dominant computational workload”, and that this is driving interest in faster alternatives to writing one token at a time, “especially as models produce increasingly long outputs” [1].
If text can be written in a few parallel steps instead of hundreds of one-by-one steps, answers could arrive faster and cost less to produce. That is the promise. It has not yet been shown at the scale of commercial chatbots.
The caveats
- Small, early research. The tests use standard research datasets and small models, not large commercial systems.
- Compared mainly with diffusion models. The speed and quality claims are against diffusion language models, which are themselves experimental. Apart from Sudoku, NCM was not tested against autoregressive models, so this does not show it beats the models behind today’s chatbots.
- Not independently checked. All the figures come from the authors. TSN could not read the paper’s full text, and the code repository linked from the blog was not publicly reachable when TSN checked.
The authors put it this way: “We are excited that such a different way of generating text works at all, and produces strong early results” [1]. Next, they want to test how NCM scales with more parameters, compute and data, train its two stages together, and let it spend more effort on harder examples [1].
Who made it
The work is led by Hermann Kumbong and Ishan S. Khare, who contributed equally, with Iddah Mlauzi, Andrew Woen, Charles Ding and E. Kelly Buchanan. Simran Arora, Azalia Mirhoseini and Christopher Ré advised equally [1].
Sources
- Hazy Research (Stanford), “Next-Concept Modeling: Fast, Parallel Coarse-to-Fine Text Generation”, 12 October 2026. https://hazyresearch.stanford.edu/blog/2026-10-12-ncm
- Paper on alphaXiv, “NCM: Next-Concept Modeling: Fast, Parallel Coarse-to-Fine Text Generation”. https://www.alphaxiv.org/pdf/2610.next-concept-modeling-parallel-generation
