The archive · AI & Models · Strategic decision · 2022–2026
ModelBest bets edge LLMs beat the cloud; Moutai-funded startup hits unicorn in 2026
Tsinghua NLP spinout ModelBest bets on-device models beat cloud LLMs; Moutai backed it in 2025, and 2026 rounds pushed it past $1B.
ModelBest (面壁智能)
What the business is
ModelBest (面壁智能) builds efficient, open-source large language models — the MiniCPM 'pocket cannon' family — designed to run fully on devices (phones, cars, robots) rather than in cloud data centers, and sells deployment with chip and automaker partners.
How it started
Founded in 2022 by researchers from Tsinghua University's natural-language-processing lab, with associate professor Liu Zhiyuan as co-founder and chief scientist; ex-Zhihu executive Li Dahai became CEO in 2023. The founding bet was that China's model race would not be won by the biggest cloud model alone: devices would need small, high-density models that run offline.
What happened
MiniCPM series downloads passed 10 million after the January 2025 release of MiniCPM-o 2.6, an 8B full-modal on-device model. In May 2025 a new round co-invested by Hongtai Fund, Guozhong Capital, Qingkong Jinxin and Moutai Fund — an unusual AI bet for the liquor giant's fund — was read as proof the edge route stayed investable after DeepSeek chilled foundation-model funding. Commercialization followed: the cpmGO car assistant, partnerships with Qualcomm, Intel and MediaTek, the April 2025 Changan Mazda EZ-60 with an on-device model, and robotics demos with Booster Robotics. In 2026 a China Telecom-led round was followed by an April round co-led by Shenzhen Capital Group and Inovance Capital, pushing first-quarter funding past ¥1 billion.
How it ended up
Unicorn and scaling: as of April 2026 ModelBest had crossed the $1B valuation bar, was named a 2026 China unicorn at the Zhongguancun Forum, and was pushing edge models into cars, robots and phones.
Background
ModelBest (面壁智能) is a Chinese AI startup founded in 2022 by researchers from Tsinghua University's natural-language-processing lab, with associate professor Liu Zhiyuan as co-founder and chief scientist and former Zhihu executive Li Dahai as CEO since 2023. Its strategy is contrarian for the foundation-model era: instead of pursuing ever-larger cloud models, it builds small, high-density models that run entirely on devices — the MiniCPM family it markets as 'pocket cannons'.
The thesis, which the company calls the density law, is that phones, cars and robots cannot depend on cloud round-trips for cost, latency or privacy reasons, so a 4B-parameter model that beats GPT-3.5 on many benchmarks can win mass deployment. In January 2025 it released MiniCPM-o 2.6, an 8B full-modal on-device model, and cumulative MiniCPM downloads passed 10 million. Commercial launches followed: the cpmGO car assistant running on car-grade SoCs, the April 2025 Changan Mazda EZ-60 with an on-device model, and robot demos with Booster Robotics.
Funding continued when the Chinese foundation-model market froze after DeepSeek's rise. In May 2025 a round co-invested by Hongtai Fund, Guozhong Capital, Qingkong Jinxin and Moutai Fund (茅台基金) — an unusual AI investment for the liquor giant's fund — was widely interpreted as evidence that the edge route remained investable. In 2026 a China Telecom-led round was followed by an April round co-led by Shenzhen Capital Group (深创投) and Inovance Capital, taking first-quarter funding past ¥1 billion.
After the April 2026 round, ModelBest's post-money valuation crossed the $1 billion unicorn threshold and it was named a 2026 China unicorn at the Zhongguancun Forum's global unicorn conference. The company says its long-term claim is that 'density plus open source' gives every terminal a path to AGI; the unresolved question is whether on-device models can keep pace as frontier benchmarks advance.
What has to be true
- After DeepSeek, Chinese foundation-model funding froze — but edge models kept raising because they sell a route to revenue instead of a race to AGI.
- Density is a defensible wedge: compressing GPT-3.5-level ability into 4B parameters requires algorithmic and data work that cloud-scale players don't optimize for.
- Distribution beats benchmarks: cars (Changan Mazda), robots (Booster) and SoC partners (Qualcomm, Intel, MediaTek) create repeatable unit economics.
- The investor mix — Moutai Fund, Loongson VC, CICC Porsche, China Telecom, SCGC — shows state-linked and strategic capital underwriting the edge thesis.
What can be applied
When rivals race on scale, win on density: small offline models with distribution into cars and phones kept funding flowing through a chill that froze cloud raises.
Aftermath
As of April 2026, ModelBest had raised three rounds in the past year — including Moutai Fund, Loongson VC, CICC Porsche and Jing Guorui, on top of earlier Zhipu, Zhihu, Hubble and Beijing AI Fund backing — crossed the unicorn valuation bar, and was deploying MiniCPM in the Changan Mazda EZ-60, robotics demos with Booster Robotics, and phone/SoC partnerships with Qualcomm, Intel and MediaTek. The open question is whether on-device models keep matching cloud quality as benchmarks move.
Sources
- 茅台基金，投了“清华系”大模型企业 (Moutai Fund invests in Tsinghua-linked ModelBest)
- 面壁智能完成新一轮数亿元融资，深创投入场 (ModelBest closes new round; SCGC enters)
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card