The landscape of China's AI large models in 2025 is witnessing profound differentiation and transformation.

If user engagement serves as the measuring stick, ByteDance's Doubao currently stands out as the brightest star; in terms of overall technical prowess and global influence, DeepSeek continues to hold its ground as China's AI frontrunner; while in the realm of the most imaginative full-stack strategy, Alibaba's Tongyi QianWen unveils unique strategic advantages.

In this all-encompassing contest of technology, products, and ecosystems, ByteDance's Doubao, DeepSeek, and Alibaba's Tongyi QianWen, although once vying on the same stage, have diverged onto distinct developmental trajectories due to differing strategic choices. One delves deep into enhancing consumer experience, another prioritizes open-source technology, and the third bets on a closed-loop ecosystem.

Doubao: Capturing Hearts with 'Down-to-Earth' Charm

In August 2023, Doubao made a quiet entrance, with ByteDance simultaneously launching Cici overseas, officially stepping into China's AI large model arena.

Doubao's ascent follows ByteDance's established logic: leveraging ecological advantages for seamless integration. Super apps like Douyin and Toutiao act as natural entry points for Doubao, with AI capabilities ingeniously woven into high-frequency scenarios such as short video creation, content recommendation, and intelligent customer service, significantly lowering the barrier for user adoption.

This ecological integration strategy enabled Doubao to swiftly amass a large user base. According to QuestMobile data, among China's top three AI apps by monthly active users in September this year, Doubao, DeepSeek, and Tencent Yuanbao secured the first, second, and third positions, respectively, with MAUs of 172 million, 145 million, and 32.86 million.

Technologically, Doubao employs an MoE architecture and UltraMem sparse model, effectively managing computational costs while ensuring rapid response times. Its core strength lies not in model scale or technical flair but in accurately addressing ordinary users' desires for simplicity, speed, enjoyment, and low barriers to entry.

For young users, Doubao transcends being a mere tool; it becomes a creative expression companion. They can utilize it to 'stylishly' edit photos, effortlessly become meme creators, and leverage a rich voice library to meet personalized communication needs. For instance, treating Doubao as a classmate to vent about the challenges of studying together or as a confidant to share inner thoughts...

Notably, on December 1st, the Doubao team unveiled a technical preview version of a mobile assistant and collaborated with ZTE to release the first engineering prototype, the nubia M153, priced at 3499 yuan. Users can directly activate Doubao through various methods and inquire about screen content in any interface, further bridging the gap between AI and users.

However, Doubao's 'down-to-earth' approach also harbors hidden concerns. When users perceive Doubao as a free tool, converting them into paying customers becomes challenging; when application scenarios are limited to content creation and entertainment, expanding into high-value fields like finance, healthcare, and education encounters obstacles, revealing slight weaknesses in high-value industry adoption.

As AI competition intensifies, mere 'ease of use and fun' is insufficient to build long-term competitive barriers. Doubao needs to discover effective pathways to penetrate the B2B market while maintaining its consumer edge.

DeepSeek: Open-Source Pioneer, Shaping AI Influence

If Doubao excels in user scale, DeepSeek triumphs in technical influence.

In early 2025, DeepSeek announced the official launch and full open-sourcing of its V3 version. Multiple benchmark tests demonstrate that DeepSeek-V3 not only surpasses mainstream open-source models like Qwen2.5-72B and Llama-3.1-405B but also matches the performance of GPT-4o and Claude-3.5-Sonnet, breaking the Western monopoly in AI technology.

On December 2nd, DeepSeek unveiled a new generation of models, DeepSeek V3.2 and DeepSeek-V3.2-Speciale. The former performs comparably to GPT-5 in multiple tests, while the latter's high-performance version achieves reasoning capabilities close to Google Gemini 3 Pro.

At the underlying technology level, DeepSeek synthesized large-scale training data by constructing virtual environments, encompassing 24,667 real code environment tasks, 50,275 real search tasks, and tens of thousands of artificially synthesized scenarios. This deep training enables the V3.2 series to achieve profound integration of thinking processes and tool invocation, significantly enhancing agent capabilities.

If technological innovation is DeepSeek's spear, then open-sourcing is its shield and banner. Since its inception, DeepSeek has fully opened its model weights, training data (de-identified), and inference frameworks, quickly becoming one of the most popular Chinese foundational models on GitHub.

This is not merely 'giving away models for free' but a meticulously designed ecological strategy.

Firstly, DeepSeek's open-sourcing lowers barriers, enabling developers to rapidly build vertical applications without starting from scratch. Solutions based on DeepSeek have emerged in professional fields like medical imaging diagnosis, financial fraud detection, and industrial predictive maintenance.

Secondly, as giants like Baidu, Alibaba, and NetEase integrate their national-level products with DeepSeek's foundation, a more diverse and healthy AI ecosystem is taking shape. This combination of 'giant endorsement + open ecosystem' has also rapidly elevated DeepSeek's influence.

Thirdly, at a critical juncture for global AI governance rule-making, DeepSeek's open-source strategy breaks the Western monopoly on core technologies. It demonstrates to the world that China can not only create world-class AI models but also advance technology development in a more open and collaborative manner.

In the AI large model arena, the true victor may not be the company with the most users but the one defining the technological ecosystem. DeepSeek is advancing steadily toward this goal.

QianWen: Full-Stack Strategy, Betting on an Intelligent Future

Compared to native AI applications like Doubao and DeepSeek, Alibaba's foray into AI application layout has been relatively measured. However, current mainstream consumer-facing AI products generally lack depth, coherence, and comprehensive scenario coverage for solving user problems, providing a window of opportunity for Alibaba's QianWen to catch up.

QianWen has chosen a distinct path: deeply integrating into Alibaba's ecosystem to become a 'chatty and capable' practical AI.

On November 17, 2025, the public beta version of the QianWen app officially launched, featuring Qwen3, dubbed the 'world's first open-source model,' directly competing with ChatGPT. Remarkably, within just one week of public beta, downloads surpassed 10 million, outperforming ChatGPT and DeepSeek's growth during the same period, showcasing strong market appeal.

QianWen's confidence stems from Alibaba's profound accumulation in AI underlying technology. Since fully open-sourcing in 2023, the Qwen series models have achieved a global download milestone of over 600 million, surpassing competitors like Llama and DeepSeek in performance and application breadth, becoming one of the world's most popular open-source large models.

Alibaba's flagship model, Qwen3-Max, even outperforms international top models like GPT-5 and Claude Opus 4 in multiple benchmark tests, ranking among the global top three. Behind this lies Alibaba's ambitious AI infrastructure investment plan. In February 2025, Alibaba announced a 380 billion yuan investment in AI infrastructure, aiming to expand cloud data center energy consumption tenfold by 2032.

QianWen's true core competitiveness lies in Alibaba's unique business ecosystem layout. On the consumer side, supported by Taobao and Tmall's shopping scenarios, DingTalk's office scenarios, Fliggy's travel scenarios, and Gaode Maps' navigation scenarios, QianWen possesses China's most complete internet lifestyle service matrix.

Imagine, in the future, simply telling QianWen, 'Help me order the fastest lunch delivery nearby,' and it automatically places the order on Taobao Flash Purchase; saying, 'I want to vacation in Hainan this weekend,' and it filters flights and hotels on Fliggy; or during travel, saying, 'Navigate to the office avoiding highways,' and Gaode Maps instantly plans the optimal route.

In short, QianWen's 'technology + scenarios + ecosystem' all-encompassing model, once forming a positive cycle, will establish industry barriers far exceeding any single technological advantage. In this AI marathon, QianWen is leveraging ecological advantages to compensate for its late start, laying a solid foundation for future intelligent competition.

Summary

In 2025, AI industry competition has evolved from early parameter comparisons and speed races to a comprehensive strength contest encompassing scenario coverage, ecosystem construction, and commercial validation.

In the consumer market, ultimate user experience and traffic operation have become pivotal, with Doubao leveraging ByteDance's product DNA to seize the initiative; in the technological sovereignty dimension, DeepSeek has earned respect from the global developer community with its open-source strategy; and at the application ecosystem level, Tongyi QianWen has achieved vertical breakthroughs by leveraging Alibaba's rich business scenarios.

Looking ahead, with the rise of new paradigms like embodied intelligence, multimodal reasoning, and AI Agents, this differentiation may intensify. However, one thing is certain: China's AI competitive landscape is no longer a zero-sum game but a collective effort driving the entire industry from 'usable' to 'user-friendly,' ultimately towards indispensability.