Doubao Work Opens Third-Party Models Such as Kimi, GLM, and DeepSeek to Enterprise Customers

The multi-model camp for office Agents has gained another member.

On October 29, ByteDance's Doubao Work announced that it would open third-party large model services to enterprise customers, with the first batch including models such as Kimi 3.1 Pro, GLM-5.3, and DeepSeek V4.1 Flash.

It should be noted that this opening is only available to Team Subscription and Enterprise Subscription customers; the regular individual version is not yet supported. Enterprises do not need to separately purchase APIs from model vendors and can directly use their existing AI credits to call different models, with administrators configuring model usage permissions by member, department, and group.

Previously, ByteDance's TRAE Work had already taken the multi-model route. Doubao Work's further opening of third-party models to enterprise customers this time means that ByteDance is also beginning to extend from relying on self-developed models to multi-model aggregation in the office Agent field.

This route is converging with Tencent's. Tencent WorkBuddy has already integrated models such as Hunyuan, GLM, MiniMax, Kimi, and DeepSeek.

By contrast, Alibaba's Qianwen Office is currently a relatively special case. As of now, the official model selector in Qianwen Office still mainly offers options such as Advanced, Basic, Economy, and Qwen3.8-Max, and has not yet clearly opened up the choice of third-party models.

The difference between proprietary models and multiple models is more than just offering one more choice.

As the model pool expands, office Agents face more complex technical challenges: which model should be called for the same task, and how can different models work together within a single workflow?

This in effect pushes competition toward the Harness beyond the models.

As a complete Agent execution system built around large models, the Harness covers aspects such as task decomposition, context management, tool invocation, task state preservation, model selection, and failure recovery.

Models provide reasoning capabilities, while the Harness is responsible for organizing these capabilities into workflows that can actually be executed.

When an Agent uses only one model, the platform mainly designs execution flows around that model's capabilities and characteristics. But as the model pool continues to expand, differences may exist in context length, tool invocation capabilities, output formats, response speed, and cost across different models, and the situations the Harness needs to adapt to also increase.

This means that multiple models do not naturally equal a better Agent experience. The more models are integrated, the more the platform needs the ability to coordinate different models and organize them into a stable, coherent execution chain.

Model routing is also an important part of this.

Large model applications always face trade-offs among effectiveness, speed, and cost. Simple document organization may not require calling the strongest model, while complex data analysis and coding tasks may require stronger reasoning capabilities.

The platform needs to select appropriate models based on task difficulty, timeliness requirements, and invocation cost, and when necessary, switch models during task execution.

However, model aggregation and automatic routing are not the same thing. At present, the model choices opened up by various products first address the question of whether users have more choices; whether accurate, stable, and economical automatic scheduling can be further achieved is another test.

Today, the office Agents of Tencent, ByteDance, and Baidu are all expanding access to third-party models, and Doubao Work has also begun to open multi-model services to enterprise customers, which has also drawn attention to Qianwen Office, which mainly relies on its own model system.

Alibaba faces a choice different from that of other vendors.

Behind Qianwen Office is Alibaba's self-developed Qwen model system, and improvements in model capabilities can directly translate into the product capabilities of the office Agent; real tasks generated by users in Qianwen Office can also provide feedback for the iteration of proprietary models and products.

If more third-party models are introduced, Qianwen Office can more quickly supplement model capabilities in different domains and expand the range of tasks the product can handle. But at the same time, the platform also needs to reconsider the division of labor between proprietary models and external models, and how different models participate in task execution.

It is worth noting that Alibaba does not insist on proprietary models in all of its Agent products. Its AI coding Agent Qoder already supports multiple model choices. In addition to Qwen, the official model selector also offers third-party models such as DeepSeek, GLM, Kimi, and MiniMax, and supports Auto intelligent routing to schedule among different models according to task requirements.

Overall, against the backdrop of the multi-model route gradually becoming an industry trend, whether Qianwen Office will continue to stick with proprietary models or open up to third-party models, and whether there will be new wavering in the future, deserves continued attention. Risk Disclosure and Disclaimer

Markets carry risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the special investment objectives, financial conditions, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular circumstances. They invest at their own risk based on this.