NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

The professor who says AI is shrinking: Liu Zhiyuan's 'Densing Law' and the MiniCPM bet

Tsinghua's Liu Zhiyuan argues model 'density' — not size — is what doubles every 100 days, and built ModelBest's MiniCPM to prove it on phones and cars.

2026-11-12 · 1136 words · NeuroAI
The professor who says AI is shrinking: Liu Zhiyuan's 'Densing Law' and the MiniCPM bet

A phone that translates a conversation without ever touching the cloud. A car cockpit that answers questions while offline on a mountain road. A voice assistant that lives on your wrist instead of a server farm.

These are the unglamorous places where one Tsinghua professor thinks the next AI fight will be decided — not in data centers, but in devices too small to hold a giant model.

Key takeaways

  • Liu Zhiyuan (刘知远) is co-founder (联合创始人) and chief scientist (首席科学家) of ModelBest (面壁智能), the Beijing lab behind the MiniCPM ("小钢炮") on-device model (端侧模型) series, and an associate professor at Tsinghua University's Department of Computer Science and Technology.
  • In 2025 his team published a paper in Nature Machine Intelligence formalizing "Densing Law" (密度定律): across 51 open-source base models, useful capability per parameter roughly doubled every 3.3–3.5 months (about 100 days) from 2023 to 2025.
  • ModelBest released MiniCPM 4.0 in June 2025 — the first system-level context-sparsification on-device model, in 8B and 0.5B versions, with a self-built inference engine the company says reaches up to 220x speedup in extreme cases.
  • The MiniCPM line debuted at CES 2025 and was named one of the Zhongguancun Forum's ten major scientific achievements in March 2025.
  • The bet: as models get denser, the same intelligence fits into phones, cars and robots, cutting cloud cost and keeping user data local.

A law named after a problem

Liu did not set out to coin a catchphrase. He set out to answer a question that bothers anyone who has tried to run a large model (大模型) on a phone: why does "make it bigger" keep winning when the device in your pocket has no room for it?

His answer, published in 2025 in Nature Machine Intelligence, reframes the whole race. Instead of tracking how capability grows with scale, he measures capability density — intelligence divided by parameter count. By surveying 51 open-source base models released between 2023 and 2025, his team found that density, not raw size, improved on a steep, regular curve: roughly a doubling every 100 days. Put plainly, to hit a given level of performance you needed about half the parameters every three-odd months.

He calls this "Densing Law" (密度定律). Where OpenAI's Scaling Law predicts that more compute and parameters yield more capability, Liu's version predicts that the same capability keeps shrinking onto cheaper, smaller hardware. It is a quieter claim than it sounds, and a more useful one for anyone building outside a hyperscaler's budget.

Why a professor started a company

ModelBest (面壁智能) exists to turn that paper into products. Liu co-founded it with Li Dahai (李大海), who runs the company as CEO, while Liu leads its science as chief scientist. The lab's flagship is MiniCPM, nicknamed "小钢炮" — "little cannon" — a family of on-device models designed to do the opposite of the industry default: get smaller and faster instead of bigger.

The 2025 release schedule shows the thesis in motion. In January, MiniCPM-o 2.6 arrived as an 8-billion-parameter full-modal model that the company says can listen, see and speak on-device; it topped GitHub's trending list for four days. In June, MiniCPM 4.0 became what ModelBest describes as the first system-level context-sparsification on-device model, shipping in 8B and 0.5B sizes with a self-developed inference engine, CPM.cu, that the company claims delivers up to 220x speedup in worst-case long-context scenarios and about 5x in normal use. By August it had open-sourced MiniCPM-V 4.5, a multimodal model it says is the first with "high-refresh" video understanding, and by September it released VoxCPM, a 0.5B voice-generation base model.

What is independently confirmed is the attention the work drew: MiniCPM was shown at CES 2025 in January and, in March, was named one of ten major scientific achievements at the Zhongguancun Forum, a state-backed technology conference.

What 100 days actually buys

A model that needs half the parameters every 100 days is a model that can keep moving down the hardware ladder — from a server, to a car, to a phone, to a pair of earbuds — without losing the ability to hold a conversation or read a document.

That matters most where the cloud is a liability: inside a car that loses signal, a factory floor with data-privacy rules, or a wearable that cannot afford latency. ModelBest's stated strategy is to put its models "where the user is," supporting mainstream chips from Qualcomm, Intel, MediaTek, Nvidia, Huawei, Apple and others, and its March 2025 "CPMGO" assistant was pitched as the first fully on-device smart-cockpit solution, followed in April by an in-car GUI agent built with Intel.

Where the argument is weak

Liu is quick to say the trend is not a straight line. In interviews he notes that core questions — better architectures, more efficient learning, what happens when training data runs out — remain unsolved, and that breakthroughs such as autonomous reinforcement learning, where a model creates its own training examples, are the real next milestone.

There is also a commercial risk baked into the thesis. If density keeps rising, the moat around on-device models gets shallower: the same open-weight releases from DeepSeek, Qwen and others that validate the trend also compete with ModelBest directly. A smaller, denser model is easier for everyone to copy. ModelBest has raised several hundred million yuan across late-2024 and 2025 rounds (an amount in the range of roughly US$30M–US$130M / HK$220M–HK$980M at current rates, since the company has not disclosed exact figures).

Honest limitations

This profile draws on ModelBest's own announcements, the 2025 Nature Machine Intelligence paper and its coverage in BJNews (新京报), Jiemian and Sina Finance. The performance figures — 220x speedup, 5x normal, GitHub trending, Hugging Face rankings — are company statements repeated by those outlets, not independently reproduced benchmarks, so I report them as claims. I could not find a public primary transcript of Liu's remarks and have paraphrased his positions from press interviews rather than quoting verbatim. The "several hundred million yuan" funding figure is a qualitative description of disclosed rounds, not a point estimate, and I have not converted it to a single dollar amount. Where a number appeared in only one source — for example, an unconfirmed report of a large IPO valuation — I dropped it. The Densing Law finding itself rests on a single peer-reviewed paper surveying 51 models; it is a real and citable result, but a trend line drawn over two years is not a guarantee.

What readers can do now

  • When evaluating on-device AI, compare capability-per-parameter and real battery or latency cost, not headline parameter counts.
  • If you build products, prototype with open on-device models (MiniCPM, Qwen, Gemma) before committing to a vendor, and measure offline performance rather than trusting demo videos.
  • Watch ModelBest's cockpit and phone deployments as the real test: durable, private, offline use matters more than benchmark leaderboards.

Related coverage

Cite this article

NeuroAI (脑机网), "The professor who says AI is shrinking: Liu Zhiyuan's 'Densing Law' and the MiniCPM bet," 12 November 2026, https://neuroai.site/a/na-people-liu-zhiyuan-densing-law

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Builders” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories