MiniCPM Models to Power Samsung Flagship Phones
Edge AI officially enters mass deployment.
According to exclusive reporting by Intelligence Emergence, Baichuan Intelligence — a leading edge-side large language model (LLM) startup — has partnered with Samsung Electronics to integrate its proprietary MiniCPM series of on-device models into upcoming Samsung flagship smartphones.
This collaboration marks a pivotal milestone: edge AI is no longer confined to demos or prototypes — it’s scaling across consumer hardware.
🌐 Regulatory Validation: Seven Major Mobile AI Services Cleared
On the same day, China’s Cyberspace Administration announced the official备案 (filing) of seven generative AI services for mobile endpoints, including:
- Apple Intelligence
- Huawei Xiaoyi LLM
- OPPO AndesGPT
- vivo BlueHeart Edge LLM
- Xiaomi HyperOS AI
- Samsung Galaxy AI
- Nubia Doudou Mobile LLM
This coordinated regulatory green light signals industry-wide readiness — and validates the technical maturity, safety, and compliance frameworks required for on-device AI at scale.
△ Source: Cyberspace Administration of China (CAC)
🚀 Strategic Shift: From In-House AI to Model-as-a-Service
Historically, Chinese OEMs built AI capabilities in-house: Huawei (Xiaoyi), OPPO (AndesGPT), vivo (BlueHeart), and Xiaomi (HyperOS AI). Now, two landmark integrations signal a structural evolution:
- Alibaba Qwen powers Apple Intelligence across iOS, iPadOS, macOS, and visionOS in mainland China — enabling native multimodal understanding & generation without app switching. (Alibaba’s stock rose >5% following confirmation.)
- Baichuan’s MiniCPM powers Samsung Galaxy AI — confirming third-party model vendors are becoming core AI infrastructure providers for global OEMs.
✅ Key implication: The mobile AI stack is maturing into a specialized, interoperable ecosystem — where chipmakers, OS vendors, model companies, and OEMs collaborate across layers.
🔍 About Baichuan Intelligence: Pioneering Efficient Edge AI
Founded in August 2022 and incubated at Tsinghua University’s NLP Lab, Baichuan Intelligence is among the earliest startups to bet decisively on edge-first AI, long before mainstream adoption.
Leadership & Milestones
- CEO Li Dahai: Former Partner & CTO of Zhihu
- Chief Scientist Prof. Liu Zhiyuan: Professor, Tsinghua CS Department
- CTO Zeng Guoyang: Core engineer behind China’s first open-source LLM — CPM-1 — trained at age 22
- 2026 H1 funding: >¥5 billion (≈$700M); valuation exceeds ¥20 billion (≈$2.8B) — making it China’s highest-valued edge-AI unicorn (Investor’s Business Daily)
💡 The “Density Law” Philosophy
In 2024, Baichuan and Tsinghua proposed the Large Model Density Law (Densing Law): intelligent capability per parameter doubles every ~3.5 months. This underpins their focus on efficiency over scale.
📈 Performance Highlights (2026)
| Model | Parameters | Key Capability | Benchmark Score |
|---|---|---|---|
| MiniCPM5-1B | 1.0B | Text foundation model | 17.9 (Artificial Analysis Intelligence Index) — outperforms larger open models |
| MiniCPM-V 4.6 | 1.3B | Multimodal (vision + text) | Runs smoothly on mobile (<6GB RAM); supports iOS, Android, HarmonyOS |
| BitCPM-CANN | Ternary-quantized | First domestic 3-value LLM for Ascend/Cambricon chips | 6× VRAM reduction in inference |
| MiniCPM-o 4.5 | 9B | Full-duplex, multimodal (voice/video/text sync) | Real-time cross-modal interaction on device |
- ✅ GitHub & Hugging Face downloads: >38 million cumulative
- ✅ Chip support: Full optimization across Qualcomm, MediaTek, Intel, NVIDIA, AMD, Rockchip, Huawei Ascend, and Cambricon
- ✅ Automotive deployment: SuperMate intelligent agent deployed in >300,000 production vehicles (Geely, SAIC Volkswagen, GAC, Mazda)
⚙️ Why Edge AI Is Harder Than It Looks
As CTO Zeng Guoyang stated in a recent 36Kr interview:
“Edge constraints are absolute — if a model won’t run, no amount of subsidy fixes it. High power draw causes heat; you can’t subsidize an ice pack. Efficiency isn’t easier — it’s harder than scaling up.”
Successful edge deployment demands co-optimization across:
– Algorithmic compression & quantization
– Hardware-aware kernel tuning
– Thermal & battery-aware scheduling
– Low-latency system integration
– Cross-OS compatibility (iOS/Android/HarmonyOS)
🔮 What’s Next for Samsung Galaxy AI?
Samsung — the world’s largest Android OEM — previously relied on Galaxy AI (in-house) and deep Gemini integration (via Google). With Baichuan’s MiniCPM now onboard, questions remain:
- Will MiniCPM handle specific subsystems (e.g., real-time translation, camera enhancement, private summarization)?
- How will it coexist with Gemini-powered features in Galaxy Unpacked 2026 (launching July 22 in London)?
- Is this the start of a broader model licensing strategy for Samsung — beyond Google?
Industry observers expect further announcements at Unpacked — potentially revealing modular AI architecture where best-in-class models serve distinct vertical tasks.
Article sourced from “Intelligence Emergence” — Author: Deng Yongyi