MiniCPM Series by Baichuan Intelligence Officially Ships on Samsung Flagship Phones, Marking the Dawn of Mass-Deployed On-Device AI

First Mass Deployment Milestone: Seven On-Device AI Services Simultaneously Filed with China’s CAC

On July 15, 2026, China’s Cyberspace Administration (CAC) announced the official filing of seven on-device generative AI services for smartphones: Apple Intelligence, Huawei Xiaoyi AI, OPPO AndesGPT, vivo Blue Heart On-Device Model, Xiaomi HyperOS AI, Samsung Galaxy AI, and Nubia Doudou Mobile AI. This marks China’s first regulatory framework for locally executed LLMs on end-user devices—and a watershed moment: on-device AI has moved beyond demos and beta features into regulated, standardized, and scalable commercial operation.

Critically, two non-OEM models appear on the list: Baichuan Intelligence’s MiniCPM series, serving as the foundational model for Samsung Galaxy AI, and Alibaba’s Qwen, powering Apple Intelligence across iOS, iPadOS, macOS, and visionOS in mainland China. This signals an emerging industry division of labor—OEMs focus on hardware integration and UX, while independent AI model companies deliver high-efficiency, cross-platform models—forming a four-layer stack: chip → model → OS → application.

Baichuan Intelligence: The Earliest Academic-Industrial Bet on On-Device AI, Built on the ‘Densing Law’

Founded in August 2022 and incubated from Tsinghua University’s NLP Lab, Baichuan Intelligence’s founding team combines academic rigor and engineering pedigree: CEO Li Dahai was former CTO of Zhihu; Chief Scientist Liu Zhiyuan is Professor at Tsinghua’s Department of Computer Science; CTO Zeng Guoyang trained China’s first open-source LLM, CPM-1, at age 22.

• In H2 2023, when most players chased parameter count in cloud-based models, Baichuan pivoted early to on-device AI and proposed the ‘Densing Law’: open-source LLM capability density doubles every ~3.5 months, enabling exponential reduction in required parameters for equivalent intelligence;
• Empirically validated in 2026: MiniCPM5-1B (1B params) scores 17.9 on the Artificial Analysis Intelligence Index—outperforming larger open-base models;
• MiniCPM-V 4.6 (1.3B params, 6GB RAM footprint) runs smoothly on mainstream mobile SoCs and supports Android, iOS, and HarmonyOS;
• Cumulative downloads exceed 38 million on GitHub and Hugging Face—the most widely adopted open-source on-device model family in China.

Full-Stack Hardware-Software Co-Optimization as Competitive Moat

On-device deployment demands more than model compression—it requires co-design across algorithm, silicon, OS, and thermal/power constraints. Baichuan has achieved:

• Universal Chip Platform Support: Full optimization for Qualcomm Snapdragon, MediaTek Dimensity, Intel Core, Rockchip RK-series, NVIDIA Jetson, and AMD Ryzen Mobile;
• Deep Domestic Chip Collaboration: Joint development with Huawei Ascend CANN and Cambricon MLU stacks; launched BitCPM-CANN—the first ternary-quantized LLM in China, cutting inference VRAM usage by ~6×;
• Cross-OS Runtime Maturity: MiniCPM-o4.5 (9B params) is a full-duplex multimodal model supporting synchronized voice/video/text I/O, achieving <800ms end-to-end latency and avg. power draw ≤1.2W on Samsung flagships;
• Hard-Constraint-First Philosophy: “On-device is hard constraint—if it doesn’t run, no amount of subsidy fixes it; if it overheats, you can’t subsidize an ice pack.” (Zeng Guoyang)

Industry Implication: A New Supply Chain Architecture Emerges

Historically, Chinese OEMs relied exclusively on self-developed on-device models (Huawei Xiaoyi, OPPO AndesGPT, etc.). Baichuan’s integration with Samsung and Alibaba’s with Apple proves that:

• Independent AI model vendors are becoming critical infrastructure providers—not via scale, but via intelligence density per watt, cross-chip portability, and production-grade deployment maturity;
• The CAC filing regime forces model vendors to embed safety, controllability, and compliance-by-design—MiniCPM series fully complies with all technical requirements under the Interim Measures for Generative AI Services;
• Per Jijie Media, Baichuan raised >¥5 billion in H1 2026 and reached a ¥20 billion valuation—the highest among China’s on-device AI unicorns—validating both its technical thesis and commercial execution.