DeepSeek, Kimi and Qwen Compared on Work Tasks: Testing Five Chinese AI Assistants

In a new video comparison, the editorial team pitted five Chinese AI assistants — Doubao, Qwen, Yuanbao, Kimi and DeepSeek — against typical work tasks where the need is not code but daily intellectual routine. The focus was on text summaries, data analysis, industry conclusions and business writing.

How the test worked

The starting point is simple: AI already handles a significant share of text work that used to fall entirely on people. So the authors skipped abstract benchmarks and chose a more practical scenario: seeing how different assistants behave on tasks resembling real requests from an editor, analyst, manager or researcher. This approach matters because users usually don't need the "smartest" bot in general, but a tool that saves time right now.

"AI can already solve most of the text tasks I run into at work."

The test included five prominent players on the Chinese market: Doubao, Qwen, Yuanbao, Kimi and DeepSeek. They were tested under identical or similar conditions, to assess not only answer quality but also how each model holds structure, handles wording and copes with tasks where the goal is not simply to continue a text but to extract value from it. The comparison is more about working fitness than about a flashy demonstration of capabilities.

Which skills were tested

The task set reflects the most common office use of AI: a user brings a large volume of text or data and wants a clear result quickly. These are the tasks where differences between models show most, because polished presentation alone is not enough — accuracy, compression, logic and the ability to keep the goal of the answer in view to the end all matter. This is the kind of work where an error is visible immediately and quickly affects the quality of the final document.

• Condense a long document into a short, clear summary
• Analyze data and pick out key figures or deviations
• Formulate industry insights from the given inputs
• Write a text in the required style with a clear structure

Testing these skills shows how well an assistant works as a layer on top of everyday processes. If a model summarizes well but gets tangled in numbers, it is hard to trust with an analytical note. If it writes confidently but cannot draw conclusions about an industry, it remains more of a draft generator. So such tests are useful not only for picking a "winner" but for understanding each tool's specialization. This is usually the level at which people decide whether to include a model in their permanent work stack.

Where the practical difference shows

The main value of this kind of comparison is that it moves AI from wow-effect mode into utility mode. For businesses and individual professionals, the question is no longer whether a model can answer questions at all. The question is whether you can hand it raw material, get a first pass quickly and then spend less time on edits. This is where real differences appear: who holds context better, who handles data more carefully, and who writes more coherent, publication-ready text.

The Chinese market has more and more such assistants, and competition is shifting from general promises to the quality of specific work. Users compare not parameter counts or marketing volume but how many steps remain after the model answers. If the text has to be rewritten entirely after the AI, its usefulness vanishes fast. If it delivers a clear structure, doesn't lose important details and saves at least half the preparation time, that is already a serious advantage in real work.

What this means

The AI assistant market is maturing: models are increasingly judged not by flashy demos but by how they behave in ordinary work tasks. For users this is a good guide: the thing to choose is not the "best AI overall" but an assistant for a specific type of work — summaries, numbers, industry analysis or draft writing. And the more often such comparisons are built around everyday routine, the easier it is to tell a convenient tool from a merely loud name.