ML engineering · 3 min read
Can NCM run on a MacBook without its GPU? I tested it against GPT-2 and Gemma 4
Next-Concept Modeling is a new way for AI to write text. I ran it on a MacBook Pro's CPU alone, next to GPT-2 and Gemma 4 12B. It runs in about 2 GB, and a fix of about 20 lines made its summaries 2.5 times faster.
Ken Ruiz Inoue · October 2026
The test
NCM, short for Next-Concept Modeling, is a new way for AI to write text, from Stanford and Together AI. Most language models write one token at a time. NCM sketches the whole text first, then sharpens all of it in a few steps.
Its paper measured speed on a data center GPU. I ran the released NCM models on a MacBook Pro's CPU alone, with its GPU unused, next to GPT-2 and Gemma 4 12B.
0
GPUs used
~2 GB
of memory
0.42 s
per news summary, after the fix
0.47 s
per short story
Yes, it runs, in about 2 GB
Both released NCM models ran on the CPU of an Apple M1 Pro alone, and the news summary model needed about 2 GB of memory for the whole process. Its news summaries scored 27.5 on ROUGE-L, a standard score for how close a summary comes to one written by a person. The paper reports 28.0, close enough that the laptop changed the speed, not the quality.
News summaries, slower than GPT-2 until one fix
NCM needs 9 steps for a summary. GPT-2, a model of about the same size that writes one token at a time, needs 20 to 28. Yet the released NCM code was slower, 1.06 s against 0.61 s, because each of its steps reread the whole 512-token article. GPT-2 reads it once and reuses that work.
The article never depends on the summary, so that work comes out the same in every step. I made NCM read it once and reuse it, with about 20 lines of code outside NCM itself. Summaries dropped to 0.42 s, faster than GPT-2, and all 100 came out word for word the same.
One news summary
Seconds on the laptop CPU, lower is better
Short stories, about 6 times faster than GPT-2
With no article to reread, the design pays off on its own. A 128-token story took NCM 11 steps and 0.47 s. GPT-2 took 128 steps and 3.01 s, and Gemma 4 12B took 16.4 s to write it.
One 128-token story
Seconds to write it, lower is better
Fast is not the same as good
Gemma 4 12B is about 80 times bigger, and it writes far better. On the same 20 articles its summaries scored 39.0 against NCM's 27.1, at about 25 seconds each on this CPU. NCM's released models are small research models, one per task, and their text often breaks down.
Here is one BBC article, summarized three ways.
A person wrote "A man has appeared in court after firearms, ammunition and cash were seized by police in Edinburgh."
NCM wrote "A man has been arrested in connection with a number of firearms ammunition and ammunition were seized in Edinburgh."
Gemma 4 12B wrote "A 26-year-old man has appeared in court after police seized firearms, ammunition, and a five-figure sum of money during searches in Edinburgh."
The point
NCM runs on a laptop CPU in about 2 GB, and its few-step design is a real speed win there once the code stops redoing work. It does not replace a model like Gemma today, because its text is still rough. The habit worth copying is to time one request, not just count steps. Here, one missing cache was 60% of the time.
Next-Concept Modeling is research by Hermann Kumbong and coauthors at Stanford University and Together AI. This is an independent test, not affiliated with or endorsed by them, Apple, Google or OpenAI. Gemma is provided by Google under the Gemma terms of use.