• Chinese OEMs are shifting from large AI models to smaller, more efficient versions (e.g. from 10B to 3B) to enhance device performance by targeting key on-device use cases.
• Focus on on-device AI processing for features like call transcription and multimodal search, ensuring user data remains private and secure.
• OEMs are embedding AI more deeply into the operating system, laying the foundation for smarter, more intuitive personal assistants.
• With the rise of DeepSeek and its light-weight AI models, the future of on-device AI deployment looks bright. However, the envisioned AI agent still needs ecosystem support.

Since 2023, Chinese Android OEMs have been actively exploring ways to fully utilize AI computing power, aiming to take the lead in the on-device AI trend and capturing customers’ mindshare. They pioneered the device-plus-cloud hybrid AI deployment model, hoping to provide customers with unique AI experiences – or at least showcase their research and development (R&D) capabilities.

In 2023, these Chinese OEMs scrambled to introduce on-device AI models with parameters reaching 7 billion or even 10 billion for their latest premium models. The market initially expected on-device AI models to continue growing in size, much like how leading global AI giants upgrade their models. However, in 2024, most Chinese OEMs took a step back and adopted a more balanced approach, downsizing on-device AI models from 10 billion to smaller versions, such as 4B or even 3B versions.

Development of On-Device AI Model Deployment 2023-2024

Note: OPPO launched OPPO Find X7 in January 2024. We put it under 2023 premium cycle.
Source: Counterpoint Research

This shift can be attributed to techniques such as pruning and distillation, which make AI models smaller while maintaining similar capabilities. Additionally, the latest flagship chipsets from Qualcomm and MediaTek claim that inference speed significantly increased in 2024. Both companies report that the NPU or APU performance in their latest chipsets has improved by over 50%. In this context, why did Chinese OEMs not deploy larger AI models on devices?

The answer lies in their pragmatic approach. Most general users do not require all-powerful on-device models to stay aligned with technological advancements. A smaller model that effectively supports high-frequency scenarios is sufficient to meet customer demands as well as save precious memory space.

At this current juncture, OEMs commonly focus on the following features:

• Call note transcription: Customers typically expect their calls to remain private.
• Multimodal AI assistant: AI assistants now process inputs beyond just voice and text.
• Multimodal search in local databases: Some users prefer not to upload their files to the cloud.
• More accurate intent detection: Users expect their behavioral data to remain secure.

These features necessitate on-device processing, as they involve significant amounts of private data. In contrast, capabilities such as video generation, image generation, and photo style transformation do not involve much sensitive data and also are beyond the capabilities of smartphone hardware in the short term. Therefore, it is more pragmatic to focus on feasible improvements that users can experience in their daily lives.

The following graphic represents our hypothesis. It illustrates AI use cases and their relationship to computing power and model size. Most of the use cases shown do not require large AI models or high computing power.

AI Use Cases Map

Source: Counterpoint Research

Furthermore, these AI-powered features are primarily integrated into the smartphone operating system as pre-installed applications. This reflects smartphone OEMs’ broader strategic judgment about the industry’s future.

Generative AI is a disruptive technology that will significantly reshape the mobile ecosystem. With its strong reasoning capabilities, it can understand and predict user intent more accurately than ever before. By integrating AI with the operating system, smartphones can become much more intelligent, bringing the vision of a true personal AI assistant closer to reality.

The recent rise of DeepSeek has provided yet another boost to the development of on-device AI agents. Its lightweight design and cost-efficient deployment have further convinced the industry that on-device AI agents are feasible. This homegrown AI leader is also accelerating market education on AI in China.

However, the transformation will not happen overnight. The path to AI-powered personal assistants requires not only stronger AI models but also a robust and supportive ecosystem. We are witnessing a story that is just beginning to unfold.