Five key trends in China's generative AI market in 2025

By Denis Depoux and Bing Li

DeepSeek is bridging the technological gap between large models worldwide, emphasizing the transformation of large models into tangible application scenarios and user value. In this article, we provide insights into 5 emerging trends within the Chinese AI agent market. These trends include: the evolution of AI agent capabilities, the transformative impact of Large Multimodal Models on AI applications, the advancement of general AI assistants alongside vertical scenario agents, the rise of agent marketplaces, and the potential for integrating AI with hardware solutions.

DeepSeek achieved a staggering 100 million user growth in just seven days, igniting a wave of mass AI adoption in the Chinese market. By February 2025, China's generative AI user base had already reached 250 million. AI applications are expanding at an unprecedented pace—within 20 days of its launch, DeepSeek hit 20 million daily active users (DAU), equivalent to 42% of ChatGPT's DAU. Meanwhile, Baidu Search integrated a full-fledged version of DeepSeek, attracting over 10 million users within just one hour of its PC release.

Now, AI Agent applications powered by large model technology are significantly reshaping the traffic distribution landscape of the internet era. The emergence of new super portals is inevitable. In this article, we present key insights and reflections on the latest trends in the AI application market, aiming to provide valuable takeaways for both foundational model providers and AI application developers.

The Five Key Trends:

• AI Agents are evolving from simple "conversational interactions" to fully autonomous "task execution loops."
• Large multimodal models enable AI applications to interact with human more empathetic.
• General AI assistants are becoming super portals, while vertical, scenario-specific Agents are entering a high-speed growth phase.
• AI agent stores represent another pathway to commercialization.
• AI applications and hardware are deeply integrating, giving rise to new types of hardware.

Trend 1: AI agents evolve from "conversational interaction" to "task completion," with more Phone Use and Computer Use applications emerging.

AI agents, represented by Manus, Zhipu AI AutoGLM, and OpenAI Operator, signal the emergence of more mature product forms. AI agents are no longer confined to chat windows but are now actively involved in real-world operations. In March 2025, the Monica team launched Manus, positioning it as the "world's first general-purpose AI agent," drawing significant industry attention. Its breakthrough lies in the autonomous task decomposition and multi-tool collaboration. Users can simply issue a command like "filter 15 resumes and generate an analysis report," and the system will automatically handle the entire workflow, including unpacking files, extracting information, and sorting. Test results show that Manus achieved a 57.7% accuracy on high-level tasks in the GAIA benchmark, with 90% lower operational costs compared to traditional solutions.

These Phone Use and Computer Use types of agents outline two emerging trends: First, the interaction mode is shifting from text-based conversation to device control, as exemplified by Zhipu AI AutoGLM, which supports cross-app operations on mobile, and Claude 3.5 Computer Use, which supports computer tasks. Second, the task granularity is moving from simple dialogue to complex workflows, as seen in Manus, which coordinates multiple agents to handle projects like job application analysis with several sub-tasks.

However, to scale up, AI agents must overcome the hallucination issue. For instance, OpenAI's Operator may generate erroneous conclusions due to data contamination in academic research; while Manus performs smoothly in demonstrations, real-world applications still experience system crashes, incomplete checkout processes, and even failure to provide checkout links.

Trend 2: The evolution of multimodal large model capabilities: AI will instantly grasp human intentions through images, text, speech, facial expressions, and gestures.

Driven by advancements in large multimodal models, multimodal capabilities are increasingly becoming a standard feature of general AI assistants. Each newly unlocked modality enriches application scenarios, enhances user experience, and ultimately boosts user adoption. Since 2024, several general AI assistants have integrated multimodal capabilities—for instance, ChatGPT introduced the GPT-4o model and GPT-4o mini, while Doubao launched an image comprehension feature in December 2024, enabling one-click generation of images with specified text.

Advancements in large multimodal models are set to enhance interactive experiences significantly. They will facilitate a more accurate understanding of user commands across multiple modes of communication, leading to interactions that are smoother and more lifelike. These models will also be adept at creating facial expressions that appear natural, coordinating speech with appropriate body gestures, recognizing and responding to a wide range of emotions sensitively, and enabling intricate interactions within mixed-reality environments. These innovative features will be key competitive advantages driving the next wave of breakthrough applications.

• AI Video Generation: One of the industry's key challenges is achieving realistic and controllable facial expressions in AI-generated characters. AI video startup Runway has introduced the Act-One feature, which specializes in facial expression modeling. With a simple camera setup, users can accurately capture an actor's performance, including eye movements and micro-expressions, and seamlessly transfer them onto characters of varying proportions. This innovation opens new possibilities for character design and animation.
• AI-Powered Social Apps: With the cross-modal understanding and generation capabilities of large multimodal models, AI-driven interactions are shifting from passive responses to active empathy. For example, Replika has introduced more lifelike avatars, enhanced AR/VR experiences, and video calling, creating immersive, contextualized conversations.

Trend 3: As general AI assistants become super portals, vertical, scenario-specific agents are expected to rise in popularity in 2025.

According to China's Top 10 AI applications ranking, general AI assistants Doubao and DeepSeek have emerged as dominant portals. As of February, the gap in monthly active users (MAU) between Doubao, DeepSeek, and the third-ranked Kimi widened to 36 million. Some new entrants, such as Tencent Yuanbao, have achieved rapid growth by integrating DeepSeek's model, delivering a smooth user experience, and leveraging aggressive marketing strategies. In contrast, established AI assistants like Wen Xiaoyan, Zhipu Qingyan, and iFlytek Spark are facing the challenge of stagnation due to slower growth, where failing to advance means falling behind.

Vertical AI applications are still in a phase of rapid expansion and market reshuffling, with the potential for applications surpassing 10 million MAUs to emerge. In China's Top 10 AI applications ranking, three spots are occupied by vertical AI applications, particularly in AI-powered social networking and video. Unlike the global rankings, however, no vertical AI application in China has yet exceeded 10 million MAUs, though their growth momentum is strong. For example, the AI video tool "Ji Meng" leverages advanced image and video generation capabilities to provide users with diverse and efficient content creation experiences. Between December last year and mid-February this year, Ji Meng’s weekly active users surged from approximately 760,000 to nearly 2 million—an almost threefold increase. Meanwhile, AI social applications such as Xingye and Maoxiang are also experiencing steady growth.

Drawing insights from global markets, niche AI applications are growing rapidly in specific verticals:

• AI × Aesthetics & Dating: Notable applications include LooksMax AI, Umax, and RIZZ, which leverage AI to help users enhance their personal appearance or optimize dating conversations.
• AI × Religion & Psychology: A representative example is Bible Chat, an interactive Bible companion designed for Christians. It offers prayer guidance, scripture reading, and daily reflections to support mental well-being. In January 2025, its monthly active users (MAU) increased by 25%, reaching 3.97 million.
• AI × Logo Design: A standout in this category is Arvin®, which streamlines the logo design process with AI, allowing users to create high-quality logos effortlessly, regardless of design experience. In January 2025, Arvin®’s MAU grew by 26%, reaching 4.73 million.
• Additionally, AI-powered fitness coaching, mental health counseling, language learning, legal assistance, and elderly care are emerging fields with significant potential for further exploration.

At the same time, leading companies are actively building AI agent development platforms, making AI agent creation accessible to a wide range of developers and users. For example, OpenAI launched GPTs, ByteDance introduced Coze, Tencent rolled out Yuanqi, and Baidu unveiled the Wenxin AgentBuilder platform. These platforms and tools offer extensive pre-integrated plugins and knowledge base resources while supporting no-code and low-code development, enabling enterprises and developers to rapidly create domain-specific "expert agents" for a broader range of mid-to-long-tail applications.

Trend 4: In addition to foreseeable monetization models like freemium, tipping, and ads, AI agent stores represent another pathway to commercialization.

The rapid development of these stores is poised to create an "App Store moment" in the era of large models. AI agent stores function as digital marketplaces where developers and creators can publish their custom AI agents, offering users a broad selection while generating new revenue streams for both creators and store operators. Drawing inspiration from Apple's App Store monetization model, AI agent stores have the potential to generate revenue through platform commissions, paid promotions, and advertising exposure. Leading general AI assistants have already begun exploring this model:

• ChatGPT launched the GPT Store, positioning it as more than just an app repository—it's a dynamic platform fostering a thriving community of innovators and users. With comprehensive features such as categorization, trending lists, and curated selections, the store now hosts nearly 3 million AI agents—surpassing Apple's App Store in the number of apps in just two months.
• Doubao introduced its AI Agent Platform, innovatively establishing a "Brand Zone" to serve as a testing ground for traditional apps transitioning into AI agents. Over ten branded AI agents, including Dongchedi, Tongcheng Travel, and Taopiaopiao, have already gone live.
• Kimi launched the Kimi+ AI Agent Store, currently offering over 20 AI agents across categories such as productivity, writing assistance, social entertainment, and everyday utilities.

However, AI agent stores are still in their early stages and face challenges such as high quantity but low quality, weak user engagement, and limited stickiness. By December 2024, the number of AI agents across leading general AI assistants had exceeded 100,000, surpassing the number of WeChat Mini Programs with more than 10,000 MAU (approx. 7,000+). Yet, despite this large volume, sustainable traffic generation remains elusive.

Some platforms, such as Tongyi, iFlytek Spark, and Tencent Yuanbao, have experienced a mismatch between agent quantity and user engagement—ranking among the top three in agent count but failing to surpass 5 million MAU. In contrast, platforms like Doubao, Kimi, and Wenxiaoyan have demonstrated a more balanced growth in both agent numbers and user adoption.

Trend 5: AI applications and hardware will deeply integrate, not only reshaping existing hardware forms but also giving rise to new types of hardware.

Reshaping Existing Hardware

The fusion of AI and hardware goes beyond simply embedding large models into existing devices. Instead, it involves a paradigm shift in human-computer interaction, where system-level AI drives terminals to evolve from passive execution to active cognition. Two key categories leading this transformation are AI smartphones and AI PCs. By 2028, global AI PC shipments are projected to reach a 70% penetration rate, while AI smartphones are expected to hit 54%.

Leading hardware manufacturers are actively implementing AI hardware strategies:

• Apple is integrating ChatGPT deeply into iOS, enhancing Siri with multimodal AI models to enable more natural responses, better understanding of vague expressions, and improved multi-turn conversations. Siri will also gain the ability to analyze screen content, summarize information, perform cross-app operations, and execute hundreds of tasks within apps.
• Honor has introduced an AI PC featuring intelligent search, document summarization, AI subtitles, and AI note-taking. In March 2025, Honor officially announced its Alpha Strategy, marking its transformation from a smartphone manufacturer into a global leader in AI-driven terminal ecosystems.
• Lenovo unveiled the ThinkBook “Codename Flip” AI PC at MWC 2025, featuring an 18.1-inch outward-folding OLED display. Powered by AI, the device enables multi-tasking, segmented workspaces, and adaptive collaboration modes, showcasing AI’s potential to revolutionize productivity tools.

AI-Driven New Hardware Categories

AI is also incubating entirely new product categories, where a hybrid "hardware purchase + subscription model" is unlocking fresh commercial opportunities.

• Next-Gen Wearable Devices: AI-powered smart glasses with multimodal capabilities, supporting real-time visual analysis, voice interaction, instant translation, teleprompter features, object recognition, and voiceprint-based payments. AI rings equipped with biometric sensors to track sleep and heart rate, paired with a monthly subscription that provides personalized health analytics and recommendations.
• Emotional Companion Robots: AI-powered pet-like robots for home environments, leveraging multimodal interaction and emotional intelligence engines. These robots can analyze voiceprint changes in real time and respond with corresponding emotional feedback, creating a lifelike companionship experience.
• AI-Powered Productivity Tools: AI voice recorders capable of speech-to-text transcription, multilingual audio-to-text conversion, real-time interpretation, and automatic meeting summaries. AI bookmarks that automatically organize reading content, upload it to an app as a digital asset, and support summarization and social sharing.
• AI-Powered Learning Assistants: AI cameras for children, integrating visual recognition and knowledge graphs to enable smart object identification, instant educational insights from photos, and bilingual narration in Chinese and English. AI spotlight systems, which assist in posture correction, focus tracking, and dynamic lighting adjustments to help children prevent myopia and spinal curvature.

As the technological gap between global large models continues to narrow, the real competitive edge will ultimately return to fundamentals. Only by transforming AI into tangible, user-centric applications and exploring sustainable, closed-loop business models can companies seize the first-mover advantage in the next wave of explosive growth.