Qwen Office Integrates Z.AI and DeepSeek, Igniting the Model Aggregation War

Office Agents have officially entered the model aggregation war.

On the evening of August 14, Qwen Office announced the launch of two models, GLM-5.3 and DeepSeek V4 Pro, which users can select directly from the "Frontier Models" tier on the product homepage.

With the previously integrated Qwen3.8-Max, Qwen Office's frontier models now cover three major model providers: Alibaba, Z.AI, and DeepSeek.

Among them, DeepSeek V4 Pro features a 1 million token context window and a maximum output of 384K tokens, with a focus on enhancing long-horizon tasks and Agent task chaining capabilities; GLM-5.3 further improves programming and long-horizon complex engineering task capabilities.

This means that following Tencent WorkBuddy, ByteDance TRAE Work, and Baidu Kuku AI, Alibaba's Qwen Office is also moving further towards multi-model aggregation.

Compared to binding only proprietary models, the logic behind the aggregation route is that the capabilities of different models can be better leveraged. Since each provider's models have different strengths in code, complex reasoning, long text, search, response speed, and invocation costs, and a complex Agent task often requires completing multiple steps consecutively.

However, as more models are integrated, another issue arises: who should actually do the work?

Currently, Qwen Office primarily relies on users to actively select models. Based on product segmentation, tasks such as complex solutions and automated workflows can choose "Frontier Models," while lightweight tasks like quick research and spreadsheet analysis can use the lower-cost "Economy Models."

The next step is for the Agent to automatically complete model selection in the background, known as "routing."

In an ideal scenario, users do not need to understand the strengths and weaknesses of every model; they simply propose a task, and the backend schedules it based on task difficulty, model capabilities, and invocation costs: simple information extraction is handed over to faster, lower-cost models, while code and complex reasoning are assigned to more capable models.

For individual users, this directly affects task completion quality, wait times, and point consumption; in enterprise scenarios, it further corresponds to computing power expenditures.

Moreover, if office Agents accumulate a sufficiently large user base, this layer may also generate new data value.

Every real task leaves feedback on model performance: which model has a higher success rate for specific tasks, how long it takes, how many tokens are consumed, and where failures are more likely to occur. Subject to privacy, authorization, and data governance requirements, this feedback can, in turn, optimize routing strategies.

Currently, this competition is still in its early stages. The industry has not yet reached a consensus on when to switch models or whether scheduling should be based on single requests or entire task sequences.

However, with Qwen Office joining the ranks of the model aggregation route, the competition among office Agents from Tencent, ByteDance, Alibaba, and Baidu is gradually aligning on the same starting line.

As the number of models increases, the next phase of competition may well be about who is smarter at "using models."