DeepSeek Harness File photo: VCG

Chinese AI startup DeepSeek said that its open-source AI agent platform DeepSeek Harness reached 1.54 million daily active users less than three months after launch, a pace of growth that Chinese industry experts said underscores the rapid rise of Chinese AI agents and a broader shift toward building out ecosystems and real-world use.

According to a statement DeepSeek posted on its X account on Sunday, DeepSeek Harness logged 1.54 million daily active users as of November 4, while daily model usage on the platform reached 92 trillion tokens. Based on that figure, usage averaged roughly 60 million tokens a day, according to estimates by the Global Times.

The surge in usage reflects a broader commercialization trend in China's AI sector, where companies are increasingly competing on product adoption now. It also suggests that the market for AI is moving beyond chatbots and benchmark performance toward improving tools that can execute tasks, plug into workflows and retain users through an expanding ecosystem of third-party applications, analysts said. 

DeepSeek Harness is an open-source framework that enables AI models to use tools and execute tasks through a modular, plug-in-based architecture. Built around an "everything is a plug-in" architecture, the framework implements core agent capabilities including models, tools, skills, sessions, sandboxes, storage, loops, and scheduling as plug-ins, allowing developers to replace or extend any layer without modifying the framework's source code.

The surge in adoption has drawn attention from foreign media too. According to a Bloomberg report, which described DeepSeek as a "Chinese AI disruptor," DeepSeek continues to challenge Anthropic PBC and OpenAI with a wide range of applications built around open-source AI technologies. 

"DeepSeek is positioning itself as an open, developer-driven agent ecosystem to challenge OpenAI and Anthropic not just on model performance and cost, but on the breadth of applications built around its AI technology," the Bloomberg report said.

DeepSeek said on its X account that the company is planning an official plug-in marketplace where developers can publish their plug-ins and users can discover them, as well as a support program for plug-in creators, offering token grants, tech discussions, and promotion of their work. The company's data showed that the number of third-party plug-ins has grown to roughly 35,000 since Harness's launch, and about 60 percent of active users have built or installed at least one.

Experts believe that as these tools and incentive policies are put into place, if DeepSeek can continue to retain its core users, its competitive focus will shift from model pricing to competing with overseas giants such as OpenAI and Anthropic over ecosystem breadth.

Meanwhile, DeepSeek's large language model (LLM) V4.1-Flash-1017 recorded 56 trillion tokens in weekly usage from November 2 to 8, up 17 percent from the previous week and ranking first on OpenRouter, according to data cited by nbd.com.cn on Monday.

LLM usage on OpenRouter reached 210 trillion tokens over the period, up 8.25 percent from the previous week. Chinese AI models accounted for 116.18 trillion tokens, an increase of 16.71 percent week-on-week, while US models recorded 13.98 trillion tokens, down 32.4 percent. Chinese AI models have maintained a higher combined weekly token volume than their US counterparts for 28 consecutive weeks.

The figures show that the vitality of China's domestic AI application ecosystem has entered the global first-tier, Tian Feng, president of the Fast Think Institute and former dean of the Intelligence Industry Research Institute at SenseTime, told the Global Times on Monday. 

As AI competition shifts from model race to ecosystem race, China's AI is forming an application ecosystem characterized by high usage, high frequency, and strong user stickiness, Tian said, adding that the DeepSeek Harness phenomenon shows that true competitiveness lies not only in model capability, but also in whether a model can be embedded into workflows and create value.

However, Tian noted that these indicators certainly cannot replace profitability, nor can they guarantee that every enterprise will succeed, but they show that China's AI sector has a continuously expanding base, a growing number of market players and large-scale computing power support.

Tian said the estimated daily token consumption per user suggests that AI is moving from "one-question, one-answer" interactions to longer-running autonomous task execution, with the platform increasingly acting like a "digital employee." 

DeepSeek is not the only Chinese player racing to embed AI into daily workflows. An industry report from QuestMobile showed that, as of September, Tencent WorkBuddy had 32.353 million monthly active users (MAUs), Baidu Dazi had 9.77 million, Alibaba Qwen Office had 7.905 million, and ByteDance's Doubao Work had 2.049 million.