DeepSeek Harness Reaches 1.54 Million DAU in Under Three Months of Launch, AI Office Market Landscape Shifts

In China's AI office market, DeepSeek has also started to grab territory.

Last night, DeepSeek Harness released a set of rather impressive figures: less than three months after launch, daily active users have already reached 1.54 million, consuming about 92 trillion tokens per day.

This result is quite a bit faster than I had expected.

After all, China's AI office market is now packed with big tech companies. Tencent has WorkBuddy, Alibaba has Qwen Office, ByteDance has Doubao Work, and Baidu has Dazi.

Compared with these products, DeepSeek Harness was launched relatively late. Although it already has two major modes, office and code, and can complete common tasks such as documents, spreadsheets, PPTs, and programming, there is still much to improve in terms of feature richness, product interface, and user experience.

English text is being prepared.

Even so, it still achieved 1.54 million daily active users.

If DeepSeek continues to improve the product and then rolls out free quotas or an Agent Plan package, is it possible for it to break into the top three or even become No. 1 in China's AI office market?

01. Less than three months after launch, 1.54 million DAU

First, let's look at the pace of DeepSeek Harness's development.

On August 13 this year, DeepSeek Harness released version v0.1, and on September 29 it officially launched the desktop client. As of November 4, daily active users had reached 1.54 million, consuming about 92 trillion tokens per day.

It should be specially noted here that DeepSeek Harness is an independent AI Agent product and is not the same thing as the familiar DeepSeek chat app.

The main DeepSeek app has a huge user base, while the 1.54 million DAU disclosed this time covers Harness users who use the official API, and cannot simply be understood as the standalone DAU of the desktop client.

So, what level does 1.54 million DAU actually represent in China's AI office market?

02. Tencent, Alibaba, Baidu: how many DAU might they have?

QuestMobile's previously released Q3 2026 AI application market report published monthly active user data for several mainstream AI office products.

As of September, Tencent WorkBuddy had 32.353 million MAU, Baidu Dazi 9.77 million, Alibaba Qwen Office 7.905 million, and ByteDance Doubao Work 2.049 million.

Tencent WorkBuddy's lead is quite obvious, with Baidu and Alibaba following closely behind.

However, these figures are all MAU, while what DeepSeek disclosed is DAU, so a direct comparison would certainly be unreasonable.

We might as well make a data projection.

Assuming a DAU/MAU ratio of 15% for these AI office products, WorkBuddy's DAU would be about 4.85 million, Baidu Dazi about 1.47 million, Qwen Office about 1.19 million, and Doubao Work about 310,000.

Here's where it gets interesting.

Under this assumption, DeepSeek Harness's 1.54 million DAU would even exceed the estimated figures for Baidu Dazi, Qwen Office, and Doubao Work, firmly ranking second.

If the DAU/MAU ratio is raised to 20%, the situation changes again.

WorkBuddy's DAU would be about 6.47 million, Baidu Dazi about 1.95 million, Qwen Office about 1.58 million, and Doubao Work about 410,000.

At that point, DeepSeek Harness would fall behind Qwen Office, but the gap between them would be only about 40,000 DAU.

Of course, the above is only a scenario projection based on public MAU data. Both 15% and 20% are assumed values, not actual statistical results. Moreover, competitors' MAU comes from September, while Harness's DAU comes from November, so the statistical scopes are not completely consistent either. Therefore, the real ranking cannot yet be determined on this basis.

But in terms of user scale, DeepSeek Harness already has the potential to challenge the leading products in China's AI office market.

And there is one more thing about it that deserves special attention.

03. DeepSeek has another disadvantage: using it costs money

A closer look at the product mechanism of DeepSeek Harness reveals that it currently still has an obvious growth threshold.

Model calls usually need to be paid for.

The Harness software itself is free and open source, but using the official models requires topping up through a DeepSeek open platform account. If users choose other models, they can also connect third-party model APIs such as Qwen, Kimi, and GLM on their own, with fees calculated according to the rules of the corresponding service providers.

Of course, if third-party free models are connected, or other model packages have already been purchased, there is also a chance to reduce usage costs.

But for ordinary users, configuring models and understanding token billing itself raises the barrier to use.

Tencent WorkBuddy already offers free trial quotas and channels for claiming points, Baidu Dazi has a daily check-in activity for claiming points, and ByteDance Doubao Work recently launched a check-in benefit as well, where 30 cumulative days of check-ins can earn 8,800 credits.

Alibaba Qwen Office likewise provides free points and some member model benefits.

DeepSeek Harness clearly still has room for improvement in this regard.

An Agent that typically requires payment for model calls and whose product experience is not yet fully polished has already accumulated 1.54 million daily active users. What would happen if DeepSeek also started giving out large-scale free quotas one day?

I think this question deserves serious consideration from Tencent, Alibaba, ByteDance, and Baidu.

For example, DeepSeek Harness could offer a certain number of tokens for daily check-ins, letting ordinary users complete a few document, spreadsheet, or PPT tasks for free.

Another example: by default, it could connect some third-party free models, using free models for simple tasks and calling more powerful paid models for complex tasks.

It could even launch a subscription package similar to Agent Plan, where users pay a fixed monthly fee and enjoy Agent task invocation services within a certain quota.

At present, DeepSeek has not announced these plans; the above models are merely conjectures about the direction of product development.

But from a product manager's perspective, these are all growth levers worth considering.

Especially given that it already has 1.54 million daily active users, if the barrier to use is lowered further and combined with continuous improvements to office features, Harness really does have a chance to win over more ordinary users.

Of course, behind free quotas lies the real cost of models; the more complex the Agent task, the more tokens it tends to consume. Whether DeepSeek can afford the subsidies, and how to convert free users into long-term paying users, are also problems that need to be solved.

03. 50,000 plugins: DeepSeek begins to pursue an ecosystem strategy

Beyond user growth, DeepSeek Harness has another advantage worth noting: its plugin ecosystem.

According to the latest official data, the number of third-party plugins for Harness has reached about 35,000, and about 60% of active users have created or installed at least one third-party plugin.

This figure is pretty impressive, showing that many users have begun to actively explore and extend Harness's capabilities, and even make plugins themselves.

DeepSeek adopts an "everything is a plugin" architecture, where developers can extend functionality according to their own needs and even replace some built-in capabilities through third-party plugins.

This time DeepSeek also announced that it will build an official plugin marketplace and support plugin developers through token incentives, technical exchanges, and featured recommendations.

It has a bit of the flavor of an App Store for the AI era.

DeepSeek does not have these mature office software and organizational collaboration systems, so catching up with the big tech companies will indeed be difficult.

But if more and more third-party developers are willing to make plugins for Harness, the business scenarios it can cover will become increasingly rich.

Users may even be able to simply go to the plugin marketplace and install whatever feature they need.

Of course, 35,000 plugins do not mean 35,000 mature and easy-to-use products. Plugin security, compatibility, and quality control are all issues DeepSeek must solve next.

04. DeepSeek establishes a product department and starts taking product seriously

There is one more thing I think cannot be ignored.

On October 16, DeepSeek integrated and established a product department, bringing the Harness, Web, App, Platform, and other teams under unified management, with Cui Tianyi in charge.

This person in charge previously worked at Jane Street for nine years and joined DeepSeek in March this year, where he was responsible for Harness-related business.

For DeepSeek, this is an organizational adjustment worth noting.

In the past, when people mentioned DeepSeek, the first thing they thought of was large-model capabilities. As for product experience, it never seemed to receive as much attention as the models themselves.

But no matter how strong the model is, ordinary users ultimately have to use it through a product.

Especially for AI Agents, which involve file operations, tool calls, task execution, and workflow management, the importance of product experience will only continue to rise.

Now that DeepSeek has set up a dedicated product department and brought Harness under unified management, there should be more moves in terms of feature richness, interface interaction, model integration, and the plugin ecosystem.

Image from the DeepSeek official WeChat account

In this official announcement, DeepSeek even proposed that it hopes to build Harness into the operating system of the Agent era.

That is indeed an ambitious goal.

If, in the future, users can install various Agent plugins the way they install apps and complete office work, programming, data analysis, design, and even enterprise business processes through a single entry point, then Harness's imaginative space would go far beyond being just an AI office software.

English text is being prepared.

Tencent has WeCom, Alibaba has DingTalk, and ByteDance has Feishu. These big companies are not short of office entry points, nor are they short of funds and technical teams.

If DeepSeek wants to enter ordinary users' daily work flows, it still needs to continuously improve the product experience so that users are willing to open it every day, not just try it out of curiosity.

But an AI Agent that has been on the market for less than three months and whose model calls usually still require payment has already reached 1.54 million daily active users. This growth rate is already enough to make other vendors tremble.

If DeepSeek next fills in the product experience gaps and then launches free quotas or an Agent Plan package, competition in the domestic AI office market will probably become even more intense.

As for whether DeepSeek Harness can break into the top three, there may be little suspense left.

This article comes from the WeChat public account "Everyone Is a Product Manager" (ID: woshipm), author: Guai Ge, published by 36Kr with authorization.