Why Modelbest Bets on Edge AI

面壁智能 · wechat · 2026-07-14

This article reviews the journey of Modelbest and its CTO Zeng Guoyang: from developing China's early large language model CPM-1, to pivoting away from competing head-to-head with cloud-based LLMs. Instead, they shifted their focus to edge AI, compressing models to run directly on devices like smartphones, cars, and toys.

The piece highlights their technical roadmap: using "knowledge density" to measure model capability and proposing the "model wind tunnel" to predict large-scale model performance through small-scale experiments. They also emphasize the critical role of data quality, categorizing internal data governance into five levels (L0-L4) and requiring engineers to manually inspect the data. Zeng notes that Modelbest has completed adaptations for chip platforms including Qualcomm, MediaTek, Intel, Rockchip, NVIDIA, and AMD. Furthermore, their new BitCPM-CANN series allows Huawei Ascend devices to fit models roughly 6 times larger within the same memory footprint.

The article also covers their engineering advancements, such as the full-duplex omni-modal model MiniCPM-o4.5, more natural edge-side memory based on behavioral pattern libraries, and ForgeTrain—reportedly the world's first production-grade training framework entirely written by AI. Overall, it's an in-depth interview blending corporate strategy, edge AI methodology, and engineering practice.

Original post →

More from Companies & People

Companies & People channel →