Four days after Doubao opened up to multiple models, Qwen switches back too

Deep News16:09

The multi-model camp for office agents has finally welcomed Alibaba.

On November 16, Alibaba's Qwen Office announced the addition of 10 large models, including the newly released Qwen4-Max, as well as third-party models from several vendors such as Kimi, GLM, DeepSeek, MiniMax, and Step.

After this update, both personal edition and enterprise standard edition users of Qwen Office can independently select models, deciding which model to use for their work based on task complexity, response speed, and Token cost.

This comes just four days after ByteDance's Doubao Work fully opened third-party models to individual users.

At this point, Tencent WorkBuddy, ByteDance Doubao Work, Baidu Kuku AI, and Alibaba Qwen Office have all moved toward a multi-model aggregation approach, and the model strategies of big tech firms in office agents are further converging.

For Qwen Office, this opening has been somewhat tortuous.

As early as August 14 this year, Qwen Office had announced the integration of two external models, Zhipu GLM-5.3 and DeepSeek V4 Pro, formally expanding its model selection range.

But just three days later, things changed.

On the evening of August 17, Wall Street See News ·全天候科技 found that the above two external models had disappeared from the Qwen Office product side, and the model selector returned to a state dominated by different capability tiers and Qwen3.8-Max.

Alibaba did not publicly explain the reason for the sudden removal of the two models.

Now, after about three months, Qwen Office has once again opened third-party models, and added 10 at once, further increasing the degree of openness.

This back-and-forth also reflects the choice Alibaba faces in the office agent field.

Alibaba itself is one of the major domestic large model developers. From Tongyi Qianwen to Alibaba Cloud Bailian, Alibaba has established a relatively complete model research and commercialization system.

For Qwen, Qwen Office is not only a window to showcase its self-developed model capabilities, but also an important application scenario for undertaking real user needs.

Once third-party models are opened, some tasks originally assigned to Qwen may shift to other model vendors.

But from the perspective of Qwen Office's own product competitiveness, opening external models is also necessary.

Office agents differ from traditional AI chat products. The tasks users deliver often involve multiple stages such as information search, data analysis, code writing, tool invocation, and deliverable output. Different tasks have different requirements for model reasoning ability, response speed, and Token consumption. Even if a model performs outstandingly in overall capability, it does not necessarily mean it can complete all work at the lowest cost.

Integrating more third-party models can expand the capability boundary of Qwen Office, allowing users to choose more suitable models for different tasks.

This means Alibaba needs to weigh the application demand for its self-developed models against the product competitiveness of Qwen Office.

Over the past few months, changes in the external competitive environment have indeed made this choice more urgent.

Tencent WorkBuddy had already integrated products from multiple model vendors including Hunyuan, GLM, MiniMax, Kimi, and DeepSeek.

ByteDance is also following quickly. On October 29 this year, Doubao Work took the lead in opening third-party models to team and enterprise subscription customers; on November 12, it further opened model selection to individual users and launched a quota consumption multiplier and discount mechanism.

Baidu's Kuku AI also already offers model choices including Wenxin, DeepSeek, GLM, Kimi, and MiniMax.

As competitors successively open third-party models, if Qwen Office continues to insist on mainly using self-developed models, it will need to rely on Qwen's performance in different office tasks to compete with the entire aggregated model lineup of its competitors.

Even if Alibaba has strong self-developed model capabilities, it is difficult to guarantee that Qwen can balance effectiveness, speed, and cost in all tasks.

If users turn to competitors because they need to use other models, what Qwen Office may lose is not just one model invocation, but also subsequent office tasks and usage habits.

This reopening of third-party models also means that Qwen Office has begun to place more emphasis on the platform's own competitiveness, rather than only the invocation demand for self-developed models.

This change is also reflected in specific product design. According to the official introduction, the newly added Qwen4-Max is mainly aimed at multi-step complex reasoning tasks, while models such as GLM-5.3 and DeepSeek V4 Pro can be used for analysis, judgment, and plan review, while relatively lower-cost models such as Qwen3.8-Flash, GLM-5.3-Flash, and DeepSeek V4.1 Flash are suitable for daily work such as copywriting and information extraction.

For enterprise customers, Qwen Office also provides model management permissions. Administrators can configure available models by department, user group, or member, matching different capabilities and costs to different positions.

Differences in capabilities among different models are beginning to be transformed into product choices oriented toward specific work scenarios.

However, with Alibaba joining, competition in multi-model office agents will also enter a new stage.

In the past, model capabilities and the scope of integration could still serve as differences between products; now, the model lineups available to each company are becoming increasingly similar, and simply integrating more models is already making it difficult to build long-term competitive advantages.

The new differentiation will come from scheduling and execution capabilities above the models.

From the product information disclosed this time, Qwen Office mainly emphasizes users' independent selection of models and has not set up an Auto mode; while Tencent WorkBuddy and ByteDance Doubao Work have further explored handing model selection to the platform through Auto mode.

There is an important difference between the two approaches.

Manual model selection requires users to judge task difficulty, model capability, and usage cost themselves; automatic model routing requires the platform to independently determine which model should be invoked based on task needs and cost constraints.

The latter is obviously more difficult. A complex task often needs to go through multiple stages such as information retrieval, data analysis, content generation, tool invocation, and result checking. Different stages have different requirements for model capabilities, and how to control Token consumption while ensuring execution effectiveness has become an important factor affecting product experience and commercialization efficiency.

Beyond model routing, agents also need capabilities such as task planning, context management, tool invocation, and error recovery. Together, these capabilities form the Agent Harness, which determines whether a task can be stably executed and delivered.

This also means that even if office agents from different companies use the same models, there may still be obvious differences in final task completion rate, execution speed, and cost.

As model selection gradually becomes a standard configuration for office agents, competition among vendors will shift further from which models are integrated to how models are scheduled and how tasks are completed.

Three months ago, Qwen Office was still wavering between self-development and openness; now, competition in office agents is making this choice clearer.

Protecting the invocation volume of its own models is certainly important, but preventing users' office tasks from flowing to competitors may be the more realistic consideration.

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.

Comments

We need your insight to fill this gap
Leave a comment