ModelBest Goes Unicorn, Edge AI On-Device

ModelBest just joined China’s unicorn club, betting big on on-device AI.

Chinese regulatory filings show a funding round worth several hundred million RMB, led by Shenzhen Capital Group and Inovance Capital, with participation from multiple institutional investors. An earlier round, led by China Telecom, helped push total fundraising in Q1 2026 to over RMB 1 billion (about USD 140 million). In just over a year, ModelBest has built a three-round path to the foundation-model unicorn tier, underscoring how far domestic AI ecosystems have traveled from the early “labs-and-pilots” phase.

What makes this milestone notable is not only the money, but the stack. ModelBest’s open-source MiniCPM family has surpassed 24 million cumulative downloads on platforms like GitHub and Hugging Face, a rare feat for a Chinese-origin project outside the Alibaba ecosystem. The company positions MiniCPM as a comprehensive on-device model stack—covering LLMs, multimodal and full-modality models, and even speech—making it one of the few Chinese players capable of rolling all those capabilities into devices without streaming to the cloud. In other words, this is not a single-model play; it’s an end-to-end software and deployment pipeline designed for edge environments.

The real-world traction is telling. MiniCPM has been deployed across automotive, smartphones, AI PCs, and smart-home devices, with production integration in models such as Changan Mazda EZ-60 and Geely Galaxy M9. The company also rolled out MiniCPM-o 4.5, described as the industry’s first full-duplex, full-modality model. And at Zhongguancun Forum 2026, MiniCPM-V 4.5 ran on an embodied AI robot, capturing motion at 10 frames per second without cloud reliance. This is a clear signal that edge inference—not cloud-only AI—has reached a scale where OEMs and automakers will treat “on-device intelligence” as a standard design criterion rather than a premium feature.

Hardware is catching up with software ambition. ModelBest has teased Songguo development boards and EdgeClaw Box devices slated for mid-2026, positions that hint at a vertically integrated strategy: develop the edge accelerator hardware to run the MiniCPM stack efficiently, then ship it through device partners. That approach mirrors a broader Chinese pattern: pairing funded AI software ecosystems with domestically manufactured hardware to forestall bottlenecks in supply and data localization.

For practitioners, this milestone carries concrete implications. First, on-device AI reshapes supply chain risk by reducing reliance on cloud data-center capacity, which matters for automakers and consumer electronics makers facing latency, privacy, or data-localization requirements. Second, the push demands domestic chip and accelerator ecosystems—memory, AI accelerators, and edge boards—reducing exposure to outside suppliers in critical segments. Third, an open-source, end-to-end stack invites rapid iteration and ecosystem collaboration, but also highlights governance and licensing questions that OEMs will need to manage as deployments scale. Finally, the wave of mid-2026 hardware introductions will test whether edge inference performance can keep up with cloud-scale expectations in real-world, automotive-grade environments.

If the dream plays out, ModelBest and MiniCPM will not just be software abstractions but a blueprint for an integrated domestic AI stack—one that could push more Chinese auto and device makers to standardize around a single edge-first foundation model family, with real turnover in who builds the devices that run it.