The competition for a national-level traffic portal cannot be decided in a single Spring Festival holiday.
“The red envelope wars have their respective strengths, and all participants are aggressive. A single holiday period is insufficient to determine a winner,” stated Chen Lei, a leader at a major AI model company, in a recent discussion.
During the previous Snake Year Spring Festival, DeepSeek intensified competition among China's large AI models. This recent Horse Year Spring Festival saw the competitive fires spread upwards to the application layer. Companies including Doubao, Qianwen, Yuanbao, and Baidu collectively deployed approximately 50 billion yuan in incentives to attract AI users, each vying to become the essential national traffic gateway in the new era.
Over the past year, Chinese enterprises have closely followed the pace of their American counterparts. In the fierce market competition, the technological gap in large models between China and the US has narrowed to approximately three months, while the application layer is nearly on par.
A single holiday is not enough to decide the outcome. “What the major tech firms are competing for is the national-level traffic portal,” Chen Lei emphasized.
This year's competition primarily involved four giants: Alibaba, ByteDance, Tencent, and Baidu.
Prior to the holiday, Alibaba's Qianwen announced a 30 billion yuan "Spring Festival Treat Plan," having already generated significant buzz within its app. Tencent launched a 10 billion yuan cash red envelope campaign in its Yuanbao app. Users engaging with Baidu's ERNIE Assistant via the Baidu App during the holiday could grab a share of 5 billion yuan in cash red envelopes. ByteDance did not disclose a specific monetary figure but stated it would distribute over 100,000 tech gifts powered by its Doubao model, along with cash red envelopes, during the Spring Festival Gala broadcast.
These major players possess distinct industrial ecosystems.
Alibaba benefits from its robust AI model and diverse application scenarios. Its series of open-source models not only lead in the domestic market but also compete effectively with American models internationally. Qianwen's holiday strategy began with free offers or subsidies like milk tea, leveraging Taobao's instant retail platform and the massive user bases of Youku and Amap, which provide rich application environments.
On February 15, Wu Jia, head of Qianwen's C-end business group, reported that user participation exceeded expectations, leading to actual holiday campaign expenditures far surpassing the initial 30 billion yuan budget. Furthermore, Qianwen's daily active user count is approaching that of Doubao.
ByteDance gained an early advantage with Doubao before the holiday competition intensified, and its proprietary model ranks among the top tier. Launched earlier, Doubao's pre-holiday daily active user numbers significantly surpassed those of Baidu's ERNIE Bot and Alibaba's Qianwen.
On February 17, Doubao disclosed that AI interactions on New Year's Eve alone reached 1.9 billion. Holiday-themed features like AI image generation and New Year greetings attracted substantial user engagement. The "Doubao Celebrates the New Year" activity helped users create over 50 million festive profile pictures and generate more than 100 million New Year祝福 messages.
In contrast, Tencent faces a noticeable capability gap in its underlying model. The prowess of its Hunyuan model currently cannot match that of the models from the Alibaba or ByteDance ecosystems.
“Can you point to a leading Tencent model on any major model ranking? No. Yet, they are confident because the requirements for a mass-market AI application are lower than those for a top-tier model. This is the basis for Tencent's confidence in this battle,” Chen Lei commented. “It's not necessarily true that a model not ranked among the best cannot win the broader market war.”
Baidu committed fewer resources to this campaign compared to Alibaba and Tencent. In the fight for the traffic portal during the holiday, it appeared more defensive than offensive.
Baidu's strategy differs because its ERNIE Assistant is integrated directly into the Baidu App, which boasts over 700 million monthly active users, eliminating the need for a separate download. The 5 billion yuan in red envelopes primarily aimed to incentivize a seamless transition for users to adopt AI features.
However, Baidu's challenge, according to Chen Lei, was that the holiday campaign's impact seemed muted at the public level, with little noticeable buzz among ordinary citizens.
The aggressive moves by these giants have raised the entry barrier significantly. For other companies to compete for the national portal, an investment comparable to these majors is now the minimum requirement, and this is just the beginning, likely deterring many potential entrants.
Chen Lei believes the ChatBot model space is now a contest among these four giants. “After this Spring Festival and 50 billion yuan in red envelopes, regardless of the outcome for these players, other companies focusing on the ChatBot model may find themselves marginalized.”
The battle for the super portal among Alibaba, ByteDance, Tencent, and Baidu will not be concluded in a single campaign. Chen Lei suggests this will evolve into a sustained conflict, similar to the food delivery wars, where one holiday season cannot determine the final victor.
Are AI Agents the Next Battlefield? Before the Spring Festival, NetEase's Youdao launched its AI agent, LobsterAI.
While OpenClaw, which previously took Silicon Valley by storm, operates on mobile devices, LobsterAI focuses on office scenarios and runs on desktop platforms. According to its product lead, users can select suitable large models for their tasks, such as DeepSeek V3.2 or Zhipu's GLM-5.
Recently, new AI agents have emerged in both China and the US. The tech community holds high expectations for agents, with many considering them the primary future form of AI application.
While large models and ChatBot portals are increasingly dominated by major tech firms—with ByteDance, Alibaba's Qianwen, and MiniMax also launching their own agents—this application format still offers significant opportunities for small and medium-sized enterprises.
The LobsterAI product lead pointed out that having a strong model and having a strong product are two different things.
“Manus doesn't have its own model, but its agent product is exceptionally well-made. Large companies naturally tend to bind their agents exclusively to their own models. This actually benefits smaller companies and startups, as they can leverage the strengths of various AI models,” the lead stated.
Manus, an AI agent company founded in China that later moved its headquarters to Singapore, was recently acquired by US tech giant Meta for billions of dollars.
Similarly, OpenClaw was not born within a large corporation but was the result of a very small startup team. It does not bind itself to a single model and can operate on models from Google or Anthropic.
A common criticism of China's mobile internet ecosystem is users' reluctance to pay for software and content, unlike more mature subscription models abroad, such as ChatGPT's.
However, this situation has improved in recent years.
Although the penetration rate of paying users in China is lower than abroad, it is significantly better than a decade ago. For instance, Youdao Dictionary generates nearly 400 million yuan in annual subscription revenue, while Meitu's imaging and design product business, primarily based on paid subscriptions, reported revenue of 1.35 billion yuan in the first half of 2025.
“If you already have a user base, it's relatively easier to layer AI features onto the existing product and monetize through subscriptions,” the LobsterAI product lead added.
Several challenges hinder the development of AI agents.
Technologically, agents must first achieve high accuracy and stability in task completion before achieving mass adoption. The challenges extend beyond technology to include integrating upstream and downstream industries, establishing cooperative relationships, and redistributing economic benefits within the value chain. The previous blocking of Doubao's mobile assistant by Alibaba and Tencent also highlights the difficulty of creating win-win cooperation mechanisms among major players.
Wang Wei, a China internet analyst at UBS Securities, believes the evolution of AI agents will occur in stages. Initially, agent functionalities may be added within individual apps, creating closed-loop operations. The next stage involves agents operating within a major firm's ecosystem, coordinating resources across different service types. Ultimately, a single agent could perform cross-platform operations, providing comprehensive user assistance and decision-making support. Collaboration between agents, the so-called A2A model, represents the final stage.
Chinese tech firms are already promoting internal use of agents. Most employees at MiniMax already use company-developed agents.
“Whether it's CRM for our sales team, HR for interviews, or, most importantly, programming, various teams utilize different agents,” Yan Yijun, Vice President of MiniMax, stated.
The US-China Gap in Large Models is Narrowing On New Year's Eve, Alibaba open-sourced a new generation large model.
The Qwen3.5-Plus model boasts 397 billion total parameters with only 17 billion activated, yet its performance surpasses the trillion-parameter Qwen3-Max model.
The period before the Spring Festival was a密集期 for releases of domestic models. ByteDance launched its video generation model SeeDance 2.0, Zhipu AI released the GLM-5 model, and MiniMax introduced MiniMax-M2.5. The highly anticipated DeepSeek has not yet released its latest model. This rapid iteration trend is expected to continue after the holiday.
According to the LobsterAI product lead's observations, the capability gap between Chinese and American large models has further narrowed from last Spring Festival to this one. The gap, which was about six months a year ago, has now been roughly halved.
Citing the Artificial Analysis leaderboard as an example, the lead noted that Zhipu's recently released GLM-5 model demonstrates capabilities in areas like programming that are close to Claude Opus 4.5. “The gap is essentially one generation, about 2 to 3 months. Looking at ByteDance's SeeDance 2.0, it has basically surpassed Google's Gemini.”
Regarding applications built on top of large models, the lead believes Chinese and American companies are on par, with Chinese firms even “slightly stronger in some aspects.”
Wang Wei holds a similar view, noting that the intelligence gap between China's top models is very small, and the average intelligence level of Chinese models is gradually catching up to the most advanced American models.
American large model companies pursue global leadership in intelligence, and with higher talent costs, the best models are often the most expensive. However, affordability is a critical consideration for practical model deployment. The improving affordability of Chinese models is driving their adoption. For example, MiniMax claims its M2.5 model can be used limitlessly without cost concerns; operating continuously for one hour while outputting 100 tokens per second costs only about $1, and about $0.3 for 50 tokens per second.
“We hope M2.5 can accelerate the arrival of the Agent era,” Yan Yijun expressed.
Major Chinese internet companies maintain a positive outlook on capital expenditure for 2026.
On January 26, Tencent founder Pony Ma acknowledged that the company's AI product Yuanbao and its large model efforts had initially “moved somewhat slowly.” This suggests Tencent may increase AI investments and accelerate its progress.
Alibaba previously announced a plan to invest 380 billion yuan over three years in AI infrastructure. In September 2025, Alibaba Group CEO and Alibaba Cloud Intelligence Group Chairman and CEO Wu Yongming stated at the Yunqi Conference that the company was actively advancing this investment plan and might commit additional funds beyond the initial amount.
Chinese companies appear to invest more efficiently in artificial intelligence compared to their American counterparts. That is, achieving a similar level of intelligence requires a smaller investment for Chinese model firms.
“Public information is becoming scarcer, making it difficult to fully measure the difference in investment scale,” Chen Lei noted, “but it is certainly significant.”
“If we measure the amount of model intelligence gained per unit of currency spent, the average cost-performance ratio of Chinese models remains highly competitive compared to the US. This allows Chinese models to maintain a leading global position in性价比, and we may even see more models expanding overseas in the future,” Wang Wei concluded.