I might be missing something, but one pretty major blind spot that I’m seeing in discussions of the China/US AI race is that no one seems to know about or discuss DouBao, which is ByteDance’s AI model. My sense of it[1] is that the use of it in China is ubiquitous (it’s like their answer to ChatGPT), and no one there really cares about Kimi or Deepseek.
Coverage of DouBao is almost entirely in Chinese, on Chinese websites, and it’s impossible to download in western app stores.
Considering that ByteDance has been on the forefront of algorithmic recommendation systems since before ChatGPT (consider how much more addictive TikTok has been than all previous forms of social media), it makes me somewhat doubtful of the estimates of how behind China is on AI development compared to US models? I don’t think anyone doing evals here has access to the Chinese frontier model!
Entirely from talking to my mom about her recent extended visit to China, and her telling me about how strange it was that every single person from ages 5-95 uses AI enthusiastically. And by AI she means exclusively DouBao. She wasn’t aware of any other Chinese AI firms.
Here is the official announcement (from a few months ago) for Seed2.0, the model family which is likely used in DouBao. The site has extensive benchmark results at the bottom, with comparisons to Western frontier models.
I understand that the announcement posts like to exaggerate but this is sort of insane, it’s a free personal trainer who can pay attention to your form in real time? God damn now I really want access to the Chinese AI.
Yes, especially their results on all kinds of visual understanding benchmarks are very impressive, sometimes significantly ahead of the competition. Unfortunately the model is really unknown outside of China. For foreigners, the website (https://www.doubao.com/chat/) redirects to a different chatbot called “Dola”. I’m not sure whether this is essentially the same model behind the scenes, just with different censorship perhaps.
I asked Kimi to search Chinese language sources for which Chinese models are best at coding and Kimi thinks Doubao is comparable to Kimi and DeepSeek on benchmarks but Chinese developers usually prefer DeepSeek for backend work. Doubao has an advantage for frontend/UI
I’m not confident Kimi is correct here, would appreciate Chinese native speakers with better understanding of the Chinese internet to redo this search/check Kimi’s work. (I can sort of read Chinese but it’s difficult/high-effort for me)
I do think the commercial wide availability of Doubao is impressive and I’ll be curious if they’ve managed to substantially drive inference costs down, or if they are just paying more.
Kimi summary
Chinese developer communities (CSDN, Juejin) consistently rank DeepSeek as the generalist coding champion for backend/logic tasks, but acknowledge specific weaknesses.
The Verdict From Chinese Tech Media
The consensus on Chinese developer forums (CSDN, Juejin, 36Kr) :
Choose DeepSeek if: You’re doing backend development, algorithmic challenges, data modeling, or need the absolute cheapest API costs
Choose Kimi if: You’re processing massive codebases (200K+ context) or academic papers with code
Choose Doubao if: You’re doing frontend/UI work (unbeatable visual coding), need multimodal integration, or want the Trae IDE ecosystem
Bottom line: On pure SWE-Bench, it’s essentially a three-way tie (76.8% vs 78.8% vs 73.1%), but Doubao’s visual coding capabilities and Trae integration create a significant real-world advantage for frontend developers that Western benchmarks—focused on text-based GitHub issue resolution—fail to capture.
I remember seeing mass adoption of XiaoIce a while back, which seems to be a similar thing. It’s not one of China’s flagship models, it’s a chatbot analogous to the enumerable consumer-facing models fine-tuned by American companies to have a specific voice, or fill a specific service niche (AI Dungeon, Character.AI, Amazon Rufus). It might be popular, but it’s not a frontier model serving as part of the AI race.
I might be missing something, but one pretty major blind spot that I’m seeing in discussions of the China/US AI race is that no one seems to know about or discuss DouBao, which is ByteDance’s AI model. My sense of it[1] is that the use of it in China is ubiquitous (it’s like their answer to ChatGPT), and no one there really cares about Kimi or Deepseek.
Coverage of DouBao is almost entirely in Chinese, on Chinese websites, and it’s impossible to download in western app stores.
Considering that ByteDance has been on the forefront of algorithmic recommendation systems since before ChatGPT (consider how much more addictive TikTok has been than all previous forms of social media), it makes me somewhat doubtful of the estimates of how behind China is on AI development compared to US models? I don’t think anyone doing evals here has access to the Chinese frontier model!
Entirely from talking to my mom about her recent extended visit to China, and her telling me about how strange it was that every single person from ages 5-95 uses AI enthusiastically. And by AI she means exclusively DouBao. She wasn’t aware of any other Chinese AI firms.
quick preliminary search of LW and EA forums found few enough hits that I can check all of the relevant ones manually. there’s:
this question on the EA forums by a new user with a Chinese username, which went unanswered.
one mention in one of Zvi’s Dec 2025 AI roundups, where he casually mentions a (native?) use case that I don’t think any western frontier model is capable of, which is simple enough to use that average parents can take advantage of it
one ignored linkpost of a newsletter covering Chinese AI, which mentions that DouBao exceeded 100 million users in September 2024
OpenAI had 300 million weekly active users in December 2024, I don’t know what exact metric “100 million users” refers to.
Here is the official announcement (from a few months ago) for Seed2.0, the model family which is likely used in DouBao. The site has extensive benchmark results at the bottom, with comparisons to Western frontier models.
I’ve tested Seed2.0 Pro on five benchmarks so far and it’s decent on them but not among the top models.
I understand that the announcement posts like to exaggerate but this is sort of insane, it’s a free personal trainer who can pay attention to your form in real time? God damn now I really want access to the Chinese AI.
Yes, especially their results on all kinds of visual understanding benchmarks are very impressive, sometimes significantly ahead of the competition. Unfortunately the model is really unknown outside of China. For foreigners, the website (https://www.doubao.com/chat/) redirects to a different chatbot called “Dola”. I’m not sure whether this is essentially the same model behind the scenes, just with different censorship perhaps.
I asked Kimi to search Chinese language sources for which Chinese models are best at coding and Kimi thinks Doubao is comparable to Kimi and DeepSeek on benchmarks but Chinese developers usually prefer DeepSeek for backend work. Doubao has an advantage for frontend/UI
I’m not confident Kimi is correct here, would appreciate Chinese native speakers with better understanding of the Chinese internet to redo this search/check Kimi’s work. (I can sort of read Chinese but it’s difficult/high-effort for me)
I do think the commercial wide availability of Doubao is impressive and I’ll be curious if they’ve managed to substantially drive inference costs down, or if they are just paying more.
Kimi summary
Chinese developer communities (CSDN, Juejin) consistently rank DeepSeek as the generalist coding champion for backend/logic tasks, but acknowledge specific weaknesses.
The Verdict From Chinese Tech Media
The consensus on Chinese developer forums (CSDN, Juejin, 36Kr)
:
Choose DeepSeek if: You’re doing backend development, algorithmic challenges, data modeling, or need the absolute cheapest API costs
Choose Kimi if: You’re processing massive codebases (200K+ context) or academic papers with code
Choose Doubao if: You’re doing frontend/UI work (unbeatable visual coding), need multimodal integration, or want the Trae IDE ecosystem
Bottom line: On pure SWE-Bench, it’s essentially a three-way tie (76.8% vs 78.8% vs 73.1%), but Doubao’s visual coding capabilities and Trae integration create a significant real-world advantage for frontend developers that Western benchmarks—focused on text-based GitHub issue resolution—fail to capture.
I remember seeing mass adoption of XiaoIce a while back, which seems to be a similar thing. It’s not one of China’s flagship models, it’s a chatbot analogous to the enumerable consumer-facing models fine-tuned by American companies to have a specific voice, or fill a specific service niche (AI Dungeon, Character.AI, Amazon Rufus). It might be popular, but it’s not a frontier model serving as part of the AI race.