I first tried the Doubao desktop app last September and liked what it could do, but I removed it after discovering that it had scanned through my hard drive and indexed all my PDF files without my authorization. From a data security standpoint, that was enough to make me uninstall it.
After that, I stuck to the web version, and I was consistently happy with it. In fact, it gradually replaced the other Chinese AI tools I had been using. I stopped using Baidu’s Wenxin Yiyan because both its answers and image generation felt mediocre, with nothing especially memorable about either. I also moved away from Tongyi because Alibaba split Q&A, image generation, and other functions across separate apps, which made the whole experience less convenient than it needed to be.
Recently I bought a new computer that doesn’t contain anything sensitive, so I decided to give the Doubao desktop client another try. The main reason was a feature highlighted in its promotion: the desktop app can deploy a model locally and use a local knowledge base at the same time, which makes fully offline, local question answering possible.
I had wanted to experiment with local models for a while, but there were always two problems. One was the setup itself, which tends to be troublesome. The other was that models small enough to run locally are usually limited in capability—something around 14B at most—and still far weaker than full online models. Doubao’s desktop app lowers that barrier to basically one-click installation, while also letting local and online capabilities coexist. After trying the current version, I ended up very satisfied with it, so it feels worth recommending to anyone with similar needs.
The screenshot above shows the full set of desktop features. “AI Search” is the general-purpose Q&A function most people will probably use the most. Of the listed tools, the only one I haven’t really tested is coding, since I’m not a programmer. That said, I did once use Doubao to build a WordPress plugin through natural-language interaction, and it produced the functionality I needed, which was impressive from a non-technical user’s point of view.
Overall, I’ve been quite satisfied with Doubao’s search, image generation, local document Q&A, AI reading, website reading, and AI agents. The only feature that hasn’t worked well for me is its writing assistant. The writing style it produces feels too stiff and machine-like.
AI Search
For the past half year, Doubao has been the AI assistant I use most for general question answering. I rarely use DeepSeek anymore. The main reason is simple: I don’t think the answer quality between the two is dramatically different, but Doubao feels smoother to use.
It handles ordinary knowledge questions well, but it can also digest long, complicated prompts and reorganize them into something coherent. For example, if asked about the relationship between precepts, concentration, and wisdom in Chinese Buddhist schools, how different traditions emphasize those three elements, and what their methods of cultivation are—answered through both Buddhist philosophy and theology—it can still produce a fairly complete, mostly accurate, and wide-ranging response. That is not an easy question. Most Buddhists probably wouldn’t know how to answer it in a structured way.
At this point, AI Q&A can essentially replace the traditional search engine for me. Whether it’s everyday browsing or actual study, conversational AI is simply more useful. It is hard to imagine going back to old-style search as the primary tool.
Image generation
Lately, Doubao has also been my main tool for image generation. You can describe what you want in natural language, choose an aspect ratio, and pick a style. In most cases, one or two rounds of revision are enough to get an image suitable for use on a website.
The extra tools are practical too: background removal, object erasing, outpainting, and image sharpening all come in handy in ordinary use.
While testing the desktop app, I found a feature that generates a similar image based on an existing one. One of the sample images was a zodiac ox, so I asked Doubao to generate a zodiac dog instead, and I was very happy with the quality.
It can also turn generated images into video, though that feature is currently still labeled as being in testing. I asked it to make a “Heavenly Dog barking at the sun” video from the dog image it created. The result looked more like a dog yawning than anything mythic or imposing. So yes, image-to-video still clearly feels experimental.
https://www.leitao.cn/wp-content/uploads/2025/06/生肖狗视频-3.mp4
Local document Q&A
This is the feature that mattered most to me.
What Doubao calls local file Q&A is essentially a combination of a local model and a local knowledge base. The local model download package is about 7 GB. As far as I can tell, this local model is for vectorization rather than inference. The actual reasoning model still seems to be Doubao 1.5 Pro AS1 Preview—the same model used by the web version.
The local knowledge base currently lets you select individual files, but not folders, which is something that really needs improvement. It supports common document formats, including PDF and Markdown files, and there doesn’t appear to be any obvious limit on how many files you can add. No official size limit is stated, but individual files under 100 MB seem to import without a problem, which means ordinary books and documents can be used normally.
My own ideal version of this feature would be much closer to a true personal knowledge system: say, preparing 1,000 PDF books in one folder, vectorizing them all at once, and then letting the model reason across that entire body of material. That would come a lot closer to how a human mind works after learning from the same library.
Right now, Doubao is not there yet.
Something like this may be possible to build manually in Cherry Studio, but when I tried Cherry Studio, its knowledge base never seemed to fully retrieve the material I had fed into it during inference. It would only pull part of it. I’m not sure whether that was because of the vector model I used, because Cherry Studio itself still has limitations, or because my expectations of a personal knowledge base are unrealistic.
Even so, Doubao’s local file Q&A is still useful in practice. Importing files one by one is enough for most daily use cases, and when privacy or confidentiality matters, having a low-barrier local workflow is valuable.
AI Reading
AI Reading feels like the online counterpart to local file Q&A. The notable part is that it supports both reading from links and reading from an AI cloud drive, and at the moment the cloud storage appears to be unlimited.
It can generate summaries, outlines, and answer questions based on the material. For ordinary reading support, that is already enough.
AI website reading
This feature was more useful than I expected. Its webpage summaries and video summaries both work well. It is especially helpful with articles that are difficult or highly specialized. It can still extract the structure and main line of argument from them.
Reading the summary first and then going through the original text makes hard material much easier to understand and easier to remember.
One feature here that I especially like is turning a webpage into a podcast-style conversation. Instead of merely reading the source material aloud, it creates a dialogue based on an understanding of the page, imitating the format of an interview or discussion. That makes difficult writing much easier to absorb by listening.
I had previously seen this kind of function in Google NotebookLM and was genuinely impressed by it, so it was surprising to see Doubao follow up so quickly.
For people who enjoy reading, this is extremely useful. Doubao’s current limitation is that this podcast generation only works from webpages, not from large documents. If it could turn an entire book into a multi-episode podcast series, quite a few jobs might start to look less secure.
What makes the result impressive is that it can identify the key points, guide the listener’s thinking, and even imitate the pauses and reflective rhythm of human speech. It is already getting very close to a real interview.
Here is the audio it generated from a webpage titled My Reflections on Learning Knowledge:
https://www.leitao.cn/wp-content/uploads/2025/06/我学习知识的心得.mp3
AI agents
Doubao’s AI agents include a wide range of applications, and most of them are genuinely practical.
The one getting the most use in my household right now is for my son: practicing spoken English and getting help with homework. It has more or less replaced the tutoring apps we used before.
Other agent-style tools—such as generating daily work reports, writing Xiaohongshu copy, or analyzing films—will probably also be useful for a lot of people.
Final impression
Among the domestic AI products I’ve tried, Doubao is still the one I’m most satisfied with right now. If these tools eventually move into a broadly paid model, Doubao is the one I would be most willing to pay for.
Different models will always have their own strengths in different niches, but for ordinary users—especially people who cannot afford to subscribe to everything—the realistic choice is to pay for the one that performs best overall. At the moment, for me, that one is Doubao.