--- title: "ModelBest CEO Li Dahai Bets Small Models Will Win the Ground War" date: 2026-11-05 category: Builders site: NeuroAI canonical: https://neuroai.site/a/na-people-li-dahai-modelbest language: en --- # ModelBest CEO Li Dahai Bets Small Models Will Win the Ground War > From founding Google China engineer to Zhihu CTO to ModelBest CEO, Li Dahai argues the future of AI in China arrives through devices, not data centers. The headlines belong to giants: ever larger training clusters, ever longer benchmark charts, cities' worth of electricity. Somewhere in that noise, Li Dahai is asking a quieter question with uncomfortable implications — what if most intelligence people actually need never touches a data center at all? As chief executive of ModelBest (面壁智能), he has spent three years betting his company on the answer. ## From Zhihu's technology helm to a founder's seat Li's career is a tour of China's internet buildout. He holds a master's degree in mathematics from Peking University. In 2006 he joined Google China as one of its founding software engineers, working on search at the moment when Google was still negotiating how to exist inside the Chinese market. Rather than settle into the multinational, he kept founding things. In 2010 he co-founded YunYun (云云网), an early attempt at a social search engine for Chinese users. He later joined mobile app store Wandoujia to lead its search technology, and in 2015 he joined Zhihu, China's largest question-and-answer community, going on to serve as its chief technology officer. Over nearly a decade there he built the advertising, data and recommendation systems behind one of the country's most valuable content platforms as it moved toward its dual listing. The turn came in 2023. Zhihu had just led an angel round in ModelBest, a small startup incubated by Tsinghua University's natural language processing lab, where researcher Liu Zhiyuan and colleagues were developing compact language models. In June 2023 Li left his perch to become the company's co-founder and chief executive; by mid-2024 he had stepped back from his Zhihu operating roles to focus on the startup full time. ## The density bet Most of China's frontier-model companies have competed the obvious way: more parameters, more GPUs, more capital. ModelBest made a subtraction instead. Its founding claim, which the team calls the Model Density Law (密度定律), holds that the parameter count needed to reach a given level of intelligence falls by roughly half every three and a half months. If the curve holds, the intelligence that today requires a server hall will tomorrow fit inside a phone. That bet produced MiniCPM, an open-source family of small multimodal models designed to run on-device. The series has accumulated downloads in the tens of millions, and the company reports deployments across smartphones, cars, AI PCs and smart home devices, including mass-produced vehicles with partners such as Geely. Instead of chasing every benchmark crown, the pitch is pragmatic: a small model close to the user, with low latency, low cost and no network dependency, can win tasks where a frontier model is overkill or unavailable. ## Why the ground war matters Li's argument is less about physics than about access. A model that must live in a distant data center needs someone to pay for the pipe; a model that lives in the device needs no permission at all. For automobiles, factory equipment, headphones and rural deployments, that difference decides what gets built. It also decides who owns the user relationship — the company that runs the giant model, or the company that owns the device. The strategy has attracted the right kind of attention. Through 2025 and 2026 the company raised multiple rounds, including a strategic investment led by state-owned telecom operator China Telecom in February 2026 and a further round led by Shenzhen Capital Group with industrial investor Inovance's arm in April 2026. In August 2026, ModelBest filed for an A-share IPO with CITIC Securities as sponsor — a deliberate contrast with peers that chose offshore listings, and a sign the company intends to build its next chapter inside China's domestic capital market. Li's own arc gives the bet credibility. He has lived through the era when search was the frontier, the era when communities and feeds were the frontier, and now the era when models are. Each time, the durable value went not to the flashiest technology but to whoever controlled distribution and cost. Small models are, in his framing, simply the next distribution. ## The awkward compromises Honesty requires the other side of the ledger. Small models concede ceilings: on complex reasoning and long-tail knowledge, miniatures still trail the giants, and a hybrid architecture that quietly calls a cloud model is often the only workable answer. On-device intelligence also inherits the industry's benchmark noise — lab numbers measured under generous conditions do not always survive contact with a three-year-old phone and a noisy street. There is also the strategic risk of being a component supplier. If the intelligence lives inside someone else's device, the device maker captures the customer. Li's team responds with open-source community-building and ever-tighter co-engineering with hardware partners, but the tension is real, and his willingness to name it is part of what distinguishes him from founders who promise an endgame they cannot yet demonstrate. ## Key takeaways - Li Dahai went from founding Google China engineer to YunYun (云云网) co-founder to nearly a decade as Zhihu's CTO before becoming co-founder and CEO of ModelBest (面壁智能) in June 2023. - The company's core thesis, the Model Density Law (密度定律), holds that the parameters needed for a given level of intelligence halve roughly every 3.5 months, pushing capable models onto phones, cars and home devices. - ModelBest raised rounds in 2026 backed by China Telecom and Shenzhen Capital Group, and in August 2026 began IPO preparation on the A-share market with CITIC Securities as sponsor. ## Honest limitations - MiniCPM's download figures circulate in different versions across company statements and media reports; treat any single cumulative number with caution. - Benchmark claims for small models come mostly from the company and its open-source community; independent evaluations on consumer devices remain limited. - The IPO is at the tutoring stage as of late 2026, not a completed listing, and revenue details for ModelBest remain private. ## What readers can do now - If you build apps, try running a small open model like MiniCPM on a local device before defaulting to a cloud API; latency and privacy change the product. - If you follow China's AI listings, watch how ModelBest's A-share filing progresses — it previews how domestic capital markets price on-device strategies. - If you manage an AI budget, audit how many of your workloads genuinely need a frontier model; the density argument applies to your cost line, not just to devices. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-people-li-dahai-modelbest Free to quote with attribution and a link to the original.