The Biggest Buyer of China's Open-Source Models Has Emerged

Deep News
Nov 20

Internet giants are becoming the new major customers of open-source large models.

Tencent President Liu Chiping revealed at the third-quarter earnings call that WorkBuddy has become the largest customer of nearly all open-weight models in China.

This may only be the beginning.

Recently, leading domestic office agents have been intensively adjusting their product strategies. After Tencent (00700.HK) led the way with a multi-model approach through WorkBuddy, ByteDance's Doubao Work has successively expanded its third-party model integration, and Alibaba's Qwen Office has also reopened its doors to external models.

From Tencent (00700.HK) to ByteDance and Alibaba, the office agents of several major internet companies are converging toward a similar product path.

As this trend becomes clearer, office agents with massive user bases and the ability to continuously take on work tasks are expected to become new super customers for upstream model providers.

But as the paths converge, the giants must not only compete for users and work entry points through office agents, but also control costs amid growing token consumption.

Whoever can aggregate more model demand determines business scale; whoever can dispatch these models more efficiently is beginning to shape the quality of commercialization.

A Super Buyer on the Way

WorkBuddy's ability to become the largest customer of nearly all open-weight models in China is partly related to its market influence, and partly because it adopted a multi-model strategy early on.

When it launched in March this year, WorkBuddy already supported integration with third-party models such as DeepSeek, GLM, and Kimi, and it has continued to expand its model selections since then.

This approach did not become the mainstream choice for domestic office agents for some time.

On August 25 this year, ByteDance officially launched Doubao Work. In its early days, the product mainly relied on its self-developed Doubao model and did not open third-party model options to regular users; during the same period, Alibaba's Qwen Office also focused primarily on its self-developed models.

But as WorkBuddy continued to expand its lead, competitive pressure on other giants surged.

On October 29, Doubao Work took the lead in opening third-party models to team and enterprise subscribers. On November 12, it further expanded model integration, adding third-party models such as Kimi K3.1 Pro, GLM 5.3, and DeepSeek V4.1 Flash alongside its original self-developed models.

Just four days later, Qwen Office quickly followed suit, launching 10 models at once, including the self-developed Qwen4-Max, as well as third-party models from Kimi, GLM, DeepSeek, MiniMax, Step, and others.

In fact, Alibaba had wavered on whether to open up third-party models.

In August this year, Qwen Office briefly introduced external models such as GLM and DeepSeek, but those models were taken offline just days after launch, and model selection returned to being primarily self-developed.

Three months later, Qwen Office once again opened its doors to third-party models.

Unlike the previous short-lived attempt, this time Qwen Office not only opened model selection to individual and enterprise standard-version users, but also allowed enterprise administrators to configure model usage permissions by department, user group, and employee.

From Tencent (00700.HK) being the first to adopt multi-model aggregation, to ByteDance expanding third-party model integration, to Alibaba reopening model selection, the office agents of several leading domestic giants have ultimately converged on a similar product path.

From the demand side, for office agents, what users ultimately need is a report, a spreadsheet, a piece of code, or a deliverable task.

Different tasks have different requirements for model capability, response speed, and cost. A single model may not maintain an advantage in all scenarios; by integrating multiple models, users can choose the more suitable model for different tasks.

This convergence in product paths is precisely creating new opportunities for upstream model providers.

In the past, independent model providers such as Zhipu, Moonshot AI, and MiniMax mainly expanded model usage through their own products, developer ecosystems, API services, and enterprise clients. Now, office agents with millions or even tens of millions of daily active users are also beginning to become a type of major customer worth competing for.

This shift is already reflected in some model providers' new product launch strategies.

In early October this year, WorkBuddy integrated an anonymous model called Space Bunny. It was not until MiniMax officially released M3.1 Flash on October 19 that the outside world learned that Space Bunny was an early version of this model.

According to MiniMax's disclosure, Space Bunny processed more than 200 trillion tokens cumulatively during its anonymous run.

As Tencent (00700.HK), ByteDance, and Alibaba successively move toward multi-model paths, the underlying model capabilities each can obtain are beginning to converge, and simply adding more model choices is no longer enough to form a sustained differentiated advantage.

How to integrate the capabilities of different models, improve task delivery efficiency, and control model invocation costs is becoming a more realistic competitive issue.

Major Customers Drive a Hard Bargain

For these internet giants that are becoming important customers of large models, the continuous expansion of model invocation scale also means bearing higher operating costs.

Tencent (00700.HK) management recently made it clear at an earnings call that in the future it will gradually improve unit economics by increasing the proportion of paying users, raising token spending per paying user, and through inference optimization, model customization, and improved computing efficiency.

On the one hand, office agents need to strengthen task delivery capabilities to increase users' willingness to pay and drive usage growth; on the other hand, they need to control model inference and task execution costs to avoid revenue growth being offset by continuously expanding resource consumption.

However, as invocation scale expands, internet giants may gain new bargaining chips while bearing higher costs.

A massive procurement scale also means an opportunity to negotiate prices lower than publicly listed API rates.

A person from a vendor that has already integrated its model into WorkBuddy confirmed to Wall Street News·All-Weather Technology: "In commercial cooperation for model APIs, the discount rate is determined based on the customer's invocation scale."

"Enterprise customers can obtain model API services through channels at prices lower than the official public listing, and there is some room for negotiation in the actual price," a person from a company in southern China engaged in large-model API agency and sales told All-Weather Technology.

Overseas model providers have adopted similar mechanisms.

Anthropic explicitly states in its official API pricing rules that high-volume customers can negotiate bulk discounts separately based on specific circumstances, and enterprise customers can also obtain custom pricing and other commercial terms.

This may further affect competition among internet giants in office agents.

As WorkBuddy, Doubao Work, and Qwen Office successively move toward multi-model paths, the underlying model supply each can obtain gradually converges, but the actual procurement prices paid may not be the same.

For platforms with larger model invocation scale, enormous procurement demand is expected to translate into stronger bargaining power, thereby securing lower unit invocation prices.

Once this advantage takes shape, the giants can not only reduce model usage costs, but may also compete for users through more competitive product pricing and service quotas, further expanding invocation scale and forming a positive cycle.

This is only part of the competition. Truly delivering a task also involves the number of model invocations, token consumption, and retry costs after execution failures, which also places higher demands on the Harness behind office agents.

As the operating framework of an Agent, Harness not only determines how these steps are organized, but also affects the number of model invocations, context management, and retry strategies after task failures.

Even with the same underlying model, different Harness implementations may produce significantly different resource consumption.

In September this year, ByteDance's Seed team and external researchers published the paper "HarnessDev," specifically studying how different Harness designs affect Agent task performance and operating efficiency.

In one set of experiments, the researchers uniformly used Gemini 3.1 Pro as the underlying execution model and had different Harness implementations complete the same Terminal-Bench 2.1 test.

The results showed that the Harness built with GPT-5.5 consumed an average of about 24.3 million execution tokens, while the Harness built with Claude Opus 4.8 consumed about 94.3 million on average, a difference of about 3.9 times.

The task success rates of the two experiments were 49.4% and 52.4%, respectively. Although the latter achieved a slightly higher success rate, its token consumption increased by nearly three times.

This also means that in the future, when several internet giants' office agents successively integrate more third-party models such as GLM, Kimi, and DeepSeek, having similar model supply does not mean they can complete work at the same cost.

From procurement to execution, internet giants actually need to figure out two accounts:

One is the price of purchasing model capabilities from model providers. As invocation scale expands, the giants have an opportunity to reduce unit costs through large-scale procurement;

The other is the comprehensive cost required to complete actual work. It depends not only on model pricing, but is also closely related to model routing and Harness operating efficiency.

These two cost advantages may also reinforce each other, forming a positive cycle.

For office agents with larger invocation scale, more favorable procurement prices mean the same budget can support more model invocations and product experiments. And the continuous accumulation of real task feedback helps the giants optimize model routing, task decomposition, and execution processes, accelerating Harness iteration.

As the multi-model path gradually converges, the underlying model capabilities available to the office agents of several internet giants are becoming increasingly similar, but procurement costs, Harness efficiency, and product iteration speed may still create gaps, making it difficult to declare a winner at this moment.

Disclaimer: Investing carries risk. This is not financial advice. The above content should not be regarded as an offer, recommendation, or solicitation on acquiring or disposing of any financial products, any associated discussions, comments, or posts by author or other users should not be considered as such either. It is solely for general information purpose only, which does not consider your own investment objectives, financial situations or needs. TTM assumes no responsibility or warranty for the accuracy and completeness of the information, investors should do their own research and may seek professional advice before investing.

Most Discussed

  1. 1
     
     
     
     
  2. 2
     
     
     
     
  3. 3
     
     
     
     
  4. 4
     
     
     
     
  5. 5
     
     
     
     
  6. 6
     
     
     
     
  7. 7
     
     
     
     
  8. 8
     
     
     
     
  9. 9
     
     
     
     
  10. 10