On July 15, the Cyberspace Administration published filing records for seven on-device generative AI services — Apple Intelligence, Huawei’s Xiaoyi, OPPO’s AndesGPT, vivo’s Blue LM, Xiaomi’s HyperOS AI, Samsung Galaxy AI, and Nubia’s Doubao phone model all cleared on the same day. It was the largest collective on-device AI approval in China to date.
Most readers will conclude one thing: “Apple Intelligence finally arrives in China — two years late.” True, but it misses the more important information. Hidden in the same-day list are two deals worth real attention: Apple chose Alibaba’s Qwen, and Samsung chose ModelBest’s MiniCPM.
This is not merely Apple AI running late. It is the formal opening of market-based division of labor for on-device AI.
A Structural Change Buried Under the Information Cascade
The filing itself is a compliance formality — every vendor offering generative AI in China must clear it. What is interesting is the arrangement behind each line of the list: who supplies whose model. Huawei’s Xiaoyi, OPPO’s AndesGPT, vivo’s Blue LM, and Xiaomi’s HyperOS AI are all self-developed, with model capability held in-house. Nubia runs Doubao (ByteDance) — sibling-company synergy.
Apple Intelligence and Samsung Galaxy AI are the only two cases where the phone maker does not train its own large model but cleared filing using someone else’s. Apple picked Alibaba Qwen; Samsung picked ModelBest MiniCPM. Read together, the two deals point one direction: phone makers’ AI capability is shifting from “build everything ourselves” to specialized division of labor.
This is not Apple and Samsung being “unable” to build models. Apple’s AFM 3 family is already live overseas with deep Google Gemini integration, spanning parameter scales from 3B to Cloud Pro. Samsung’s Galaxy AI keeps integrating Gemini. But in the Chinese market, both independently chose “don’t build it ourselves” — letting external model companies fill the on-device gap. That is not a capability deficiency; it is a strategic choice. The logic is not complicated: on-device AI is systems engineering across hardware, software, silicon, power, thermals, and ecosystem adaptation. Apple and Samsung could absolutely build it — but in China, the ROI math says buying beats building. And that decision is redefining the division of labor in phone AI.
On-Device AI’s Three Eras
Era one: cloud-led (2023-2024)
“Phone AI” essentially meant “a phone app calling a cloud API.” The user opened an AI app, typed, and waited for the cloud to respond; the phone’s role was displaying results. Vendors were channels, and model competitiveness lived in cloud capability and API pricing.
Era two: self-developed on-device (2024-2025)
Qualcomm and MediaTek began packing NPUs into chips, Apple shipped Apple Intelligence’s on-device architecture, and Huawei, vivo, and others landed self-trained models on phones. The era’s marker: “the model runs locally” — AI erasure, live summaries, on-device image generation, all without the network. But the era’s problem was every phone maker rebuilding the same wheel: Huawei trained its model, vivo trained its model, OPPO trained its model — different hardware platforms, different chip adaptations, different data distributions, yet nearly identical features. Massive duplicated investment, diminishing returns.
Era three: market-based division of labor (2026-)
July 15 marks the start of stage three. Phone makers no longer need to train their own models. Model companies — Alibaba Qwen, ModelBest, ByteDance Doubao — become specialized “on-device AI capability suppliers,” while phone makers handle hardware integration and experience design and model companies handle efficiency and adaptation. This does not kill the self-developed route: Huawei, vivo, and others with in-house capability will keep iterating. But the market now has a second option — external models — and external models are improving fast.
The “On-Device Supply Chain”: Three Modes
The phone-AI landscape is splitting into three modes, each vendor self-positioning: Mode A, the fully self-developed loop. Huawei, vivo, OPPO, Xiaomi — train your own models, adapt your own chips, control your own experience. End-to-end control with privacy and data closure inside your own system, at the cost of heavy investment and the risk that fast-iterating external models outrun your internal version. Mode B, core self-development plus external supplements. Apple (self-developed AFM 3 overseas, Qwen in China) and Samsung (Gemini globally, MiniCPM in China) hold core capability in-house while introducing external models for specific markets. A pragmatic compromise — preserving long-term bargaining power and technical control while plugging short-term market gaps. Mode C, pure external procurement. Nubia plus Doubao: the phone maker positions itself purely as hardware and channel, with all AI capability supplied by the model company. An asset-light strategy suited to mid-tier vendors chasing volume.
No mode is absolutely right — they depend on scale, technical reserves, and market strategy. But the key insight: Mode C’s existence means on-device AI is no longer a big-guys-only game. Any phone maker, even any IoT device vendor, can in theory acquire AI capability by procuring an on-device model. That is the real signal behind “seven filings in one day.”
The Model Companies’ New Battlefield
If the 2024-2025 model race was “who tops the leaderboards,” the 2026 battlefield is “who gets onto the phone.” Seven filings landing together means on-device AI is no longer a technical demo — it is a consumer product. ModelBest is the case study: in late 2023, while the whole industry crowded cloud-model parameter races, ModelBest went all-in on on-device and was seen as off-mainstream. Two years later its MiniCPM series proved that “going small” is no easier than “going big” — making a 1-3B model usable on phone silicon involves algorithm compression, chip adaptation, power optimization, and latency engineering as one integrated discipline. In May 2026, MiniCPM5-1B scored 17.9 on the Artificial Analysis Intelligence Index, beating larger open models, and validated ModelBest’s “model density law”: the parameters needed for equal intelligence double every 3.5 months. If the curve holds, the on-device capability ceiling rises fast.
Alibaba Qwen being chosen by Apple is the more macro signal. Alibaba is not a pure model startup — it has cloud, e-commerce, payments, and vast user data. Apple picking Qwen as its China AI base is not only about model quality: the partnership involves data compliance, infrastructure integration, and service stability — recognition of Qwen as “AI infrastructure”-grade capability. Alibaba’s stock rose over 5% on the news; the capital market’s read was clear: landing Apple is worth far more than model service fees — it is the most expensive social endorsement Alibaba AI could buy for consumer deployment.
What Different Roles Should Do
Phone-product leads: stop selling “self-developed large model” as the flagship’s sole selling point. Users do not care who trained the model — they care whether the AI features work. If an external model is better, adapts faster, and costs less, why insist on building? Run a serious self-build-versus-procure ROI analysis across R&D investment, iteration speed, and ecosystem fit. Model-company tech leads: the leaderboard race is not over, but the endgame is no longer decided by leaderboards alone. The scarce capability is “stuffing large models into small devices and having them run smoothly” — compression, chip adaptation, power, latency; each link can become a moat. If you have no phone-vendor partnership today, ask yourself: on which device does my model run? Developers and founders: on-device AI moving from demo to consumer product opens a new application entrance. In 2024 you called cloud APIs; in 2025 you embedded AI features in apps; in 2026 you can invoke AI at the phone-OS level — and the user never needs to know whose model it is. That is a deeper distribution logic than the App Store.
Three years ago the fashionable line was “every company deserves to be rebuilt by AI.” Three years later the truer answer is: every device deserves to be rebuilt by on-device AI — and July 15, 2026, is the day that process formally started.
References: 36Kr, ifanr, and the CAC announcement.
