Doubao Work has recently opened third-party models such as Kimi, GLM and DeepSeek to personal users, while adding automatic scheduling. This change matters: the competitive focus in AI-powered work is shifting from “Who has the strongest single model?” to “Who can organize models, tools and workflows to get work done?”

What this is

In the past, Doubao Work offered third-party models mainly to enterprise customers, while individual users could generally complete tasks only within its own model ecosystem. Now, users can select a model by task or let the system decide automatically. Each model also has a labeled quota-consumption multiplier, allowing users to lower usage costs during promotional periods.

More importantly, tasks can now run continuously across different models. Intermediate outputs such as profit-and-loss analysis and structural breakdowns generated at one stage can be retained, eliminating the need to enter them again. Doubao Work also integrates workflows for document analysis, web-page screenshots, video production and PPT generation, and can invoke ecosystem capabilities including Feishu, Douyin, Jimeng and Seedance.

Industry view

This sends a clear signal: model capability is becoming infrastructure, and differentiation in office Agents increasingly comes from the Harness—the engineering system that organizes models, context, tools and execution workflows. Whoever can reliably handle model scheduling, memory, tool invocation, security and cost will be best positioned to turn model capability into real productivity.

But the risks are equally direct. The more third-party models are added, the more complex the invocation chain becomes. Whether data is separately retained by different providers, how output results can be traced, and which party bears responsibility for errors when multiple subagents collaborate all remain questions that need watching. Promotional pricing for individual users is temporary, while quota consumption and actual performance may vary by task. Automatic model selection does not guarantee better results, and important materials still require human review.

Impact on regular people

For enterprise IT: Companies can combine different models by task rather than committing all work to one model. However, they need to establish cross-model data boundaries, permissions and audit standards in advance.

For individual professionals: The barriers to completing tasks such as long-document analysis, first-draft PPTs and video production continue to fall. The value will more likely emerge in review, judgment and final editing than in simply moving content from one place to another.

For the consumer market: Opening the personal edition to third-party models, together with limited-time discounts, will accelerate AI office products’ transition from novelty tools to everyday services. But free quotas and low monthly prices may only mark the start of competition; long-term retention will ultimately depend on stability and output quality.