Summary of Key Points
Leading domestic technology companies (Alibaba, Baidu, Tencent, ByteDance) are collectively shifting towards a "model aggregation" approach for their office AI assistants. Instead of relying solely on their own large models, they are integrating third-party models such as Zhispu and DeepSeek, giving users the freedom to choose. The primary goal is to encourage users to entrust real work tasks (writing reports, creating PPTs, analyzing tables, etc.) to the AI, thereby accumulating valuable data on task breakdown, tool usage, and user feedback. This data can be used to optimize the AI's "command system." However, a significant challenge arises: how to select the most suitable model for each task while balancing performance, speed, and cost—these three factors are difficult to satisfy simultaneously, and this will become the key to future competition.
Why Are Office AI Systems Starting to Accumulate Models? — The Ultimate Goal is User Adoption
In the past, the focus of large-scale models was on which model was the most powerful. But with office AI, the approach has changed: first, it's about retaining users and getting them to actually use the tools. For example, Qianwen Office has integrated Zhispu GLM-5.3 and DeepSeek V4 Pro, while Tencent WorkBuddy has incorporated multiple models including Hunyuan, Zhispu, and MiniMax. The reason is that no single model can handle all tasks efficiently; some are better at writing copy, some at analyzing data, and others at creating PPTs. A richer variety of models means a higher probability of solving users' problems, which in turn increases their willingness to use the AI. More importantly, the more users use the AI, the more "real workflow data" it collects. For instance, when users ask the AI to write a report, how does it break down the task into steps like finding information, drafting an outline, filling in content, and refining the presentation? Where do errors occur, and how do users make adjustments? This data is crucial for the AI's improvement—just like teachers correcting assignments; only by reviewing many samples can they understand where students need help and how to provide better guidance. A Baidu executive stated that public data has already been fully utilized by models, and future improvements will rely on "new knowledge generated by users," with real workflows being the source of this knowledge.
Therefore, accumulating models is not about showing off technical capabilities but about creating a cycle where users are motivated to use the AI, which in turn generates more data, leading to better AI performance and even greater user engagement (a so-called "data flywheel").
Aggregation Does Not Mean Openness Without Limits — Big Companies Still Protect Their Ecosystems
Although companies are integrating third-party models, their model pools do not include direct competitors' models. For example, Tencent WorkBuddy does not have Alibaba's Qianwen or ByteDance's Seed; ByteDance's TRAE Work does not have Tencent's Hunyuan; and Baidu's KukuAI does not have Hunyuan, Qianwen, or Seed. The reason is that big companies do not want to direct users away from their own products. While the integration is "open," there are still barriers to entry. It's like having separate sections in a mall where you can introduce other brands but prevent competitors from taking over business.
Routing: The Critical Issue of Model Selection
Integrating multiple models is just the first step; the real challenge is to select the right model for each task, which is known as "routing." There is an "impossible triangle" at play: high-performance models are often expensive (e.g., using large models for complex tasks, costing several dollars per invocation) and slow; fast models may not perform well; cheap models might not be capable of handling complex issues. The goal of routing is to find a balance among these three factors:
- Simple tasks (e.g., generating a meeting summary): use inexpensive, fast models to save costs.
- Complex tasks (e.g., analyzing annual sales data and creating a PPT): use high-performance models to ensure quality.
- Urgent tasks (e.g., needing a report framework within 10 minutes): use fast models, even if it means sacrificing some performance.
For users, this affects the outcome, waiting time, and cost. For corporate clients, the difference in costs for batch tasks can be significant—just 0.1 yuan per invocation can add up to tens of thousands of dollars when processed millions of times. Therefore, routing is not just about choosing a model but also about making economic and performance calculations.
The Ultimate Competition: Whose Command System Is the Smartest?
Models can be purchased or integrated, but the real barrier lies in how tasks are broken down, models are selected, tools are invoked, and failures are handled—this constitutes the AI's "command system." For example, when a user requests a Q3 sales report, the AI's command system must break down the task into steps like retrieving data, analyzing trends, drafting an outline, filling in content, and refining the presentation. It then selects the right models for each step (e.g., using a model optimized for data analysis, logic, or copywriting). If something goes wrong (e.g., incorrect data retrieval), the system must be able to recover automatically. The quality of this command system depends on the amount of real task data and scheduling experience accumulated by the platform. Some platforms can more accurately determine which model is more cost-effective and efficient for a given task, while others cannot.
In summary, office AI is moving from a focus on individual models to a multi-model aggregation approach to attract users and collect data. However, to succeed, companies must first solve the problem of selecting the right model for each task. The ultimate competition will be about whose command system is the most intelligent—capable of finding the best balance between performance, speed, and cost, providing a better user experience and saving money.
In One Sentence
Office AI is shifting from competing on individual models to using multiple models together to gain more users and data. But to make this work effectively, companies must first address the challenge of model selection. Ultimately, it's about whose command system is the most advanced and efficient, providing a smoother and more cost-effective user experience.