In a Beijing lab that is easy to miss from the street, a phone sits on a desk running a multimodal AI model that can describe what its camera sees and answer in a cloned voice — with no network connection. The model fits inside memory most laptops would treat as nothing.
That phone is the whole argument Liu Zhiyuan (刘知远) has been making since 2022: the future of useful AI may not live in the cloud at all.
Key takeaways
- Liu Zhiyuan, a tenured associate professor at Tsinghua University, co-founded ModelBest (面壁智能) in August 2022 to build edge-side large models (端侧大模型) instead of chasing cloud titans.
- The company's MiniCPM "small cannon" (小钢炮) series runs multimodal models on phones and cars at 1B–8B parameters under an Apache 2.0 licence.
- Core MiniCPM-V research was published in Nature Communications in July 2026; a companion "Densing Law of LLMs" paper appeared as a Nature Machine Intelligence cover article.
- In July 2025 China's cyberspace regulator listed seven on-device model filings for phones; Samsung's Galaxy AI was built on ModelBest's underlying on-device technology.
- ModelBest filed for an initial public offering in August 2026, betting that device-side AI is a distinct market from cloud computing.
A lab spinoff with a contrarian thesis
ModelBest was incubated from the Tsinghua NLP lab (THUNLP), the same group behind early Chinese pre-training work. Liu had led pre-training research since 2018, contributed to the ERNIE model in 2019 and to the "Wu Dao 2.0" (悟道2.0) trillion-parameter project in 2020 before starting the company with Zeng Guoyang (曾国洋, chief technology officer) and Li Dahai (李大海, chief executive). Where most Chinese model startups of that wave raced to build ever-larger cloud bases, Liu's team went the other way: shrink the model until it fits the device.
The bet has a name inside the lab — "density law" (密度定律) — the idea that intelligence per watt and per parameter, not raw scale, is the metric that decides whether AI reaches ordinary users. It is a direct challenge to the "bigger is better" consensus of 2023–2024.
What "edge" actually buys
MiniCPM-V is designed to run on roughly 6 GB of memory, which puts it on commodity phones, PCs and car cockpits. The team says the 8B MiniCPM-V 4.5 matches or beats models ten times its size on certain vision-and-language benchmarks; those comparisons are vendor-reported and should be read as claims, not independent verdicts. What is not in dispute is the engineering point: a model that runs offline protects privacy, removes latency and keeps working in a tunnel.
That argument has landed in products. ModelBest's models appear in Geely's Galaxy M9 (吉利银河 M9) cockpit and in robotics demos, and the company released AgentCPM, a GUI agent tuned for Qualcomm's Snapdragon platforms, in late 2025. ModelBest's FlagOS toolchain adapts MiniCPM to nine chip platforms including ARM, arguing that device diversity — not a single vendor — is the real edge-AI constraint in China. None of this required a data centre. Larger platform vendors such as Apple and Qualcomm are shipping their own on-device models, so ModelBest competes with the very chipmakers it depends on.
The Samsung proof point
The clearest external validation came from the regulator's own list. In July 2025, China's Cyberspace Administration published a batch of seven on-device model filings for phones; among the named devices, Samsung's Galaxy AI (盖乐世 AI) was built on ModelBest's underlying on-device technology, according to the filing record reported by the financial press. For a four-year-old lab spinoff, that is an unusual route to scale: not a consumer app of its own, but as the hidden layer inside someone else's hardware.
ModelBest filed for an IPO in August 2026 with CITIC as sponsor, a milestone that turns the edge-model thesis into a test the public market will score.
According to Li An, chief scientist at BrainNet (脑机网), the strategic weight on edge-side models reflects a wider shift in China from chasing parameter scale toward optimizing for on-device efficiency, where cost and latency — not benchmark leaderboards — decide adoption.
Honest limitations
Most headline performance numbers for MiniCPM come from the company or its academic partners and have not been independently audited; benchmark wins against much larger models are framed by the vendor's own harnesses. Edge models still trail the frontier cloud models on the hardest reasoning tasks, and the IPO is a filing, not a listing — timing and valuation are uncertain. The Samsung relationship is documented through a regulatory filing, not a joint press release, so the commercial depth of that tie is unclear.
What readers can do now
- Developers: pull MiniCPM from its open-source repository and benchmark it on a real phone before assuming you need a cloud API for multimodal features.
- Product teams: treat on-device models as a privacy and offline-resilience option, and test where latency actually drops for your users.
- Investors and operators: watch ModelBest's IPO process as a signal for whether "edge AI" is a separate, fundable category or a feature bundled into chips and handsets.