AI Translation Comparison

Task

  • Recently, due to work needs, I downloaded a German engineering professional document from the Internet and needed to translate it.

Model

  • A total of 5 domestic AI models were used, all using the official web version and logging in as a free user.
    • Kimi
      • Chose the latest K2 model
    • Doubao
      • There is no model selection on the official website, so I chose deep thinking: automatic
    • Tongyi Qianwen
      • Chose the latest Qwen 3 model
    • Wenxin Yiyan
      • Chose the latest Wenxin X1 Turbo model
    • DeepSeek
      • Turned off deep thinking and chose the latest V3 version
    • Tencent Yuanbao
      • Chose the Hunyuan model

Instructions

  • A relatively simple prompt was used
You are a civil engineering expert who is proficient in both Chinese and German. Please translate the attached German guide into Chinese, output it in markdown format, use professional terms for shield engineering and civil engineering, and adopt a rigorous academic style. For unclear professional terms, please query professional academic websites. Please translate Chapter 1 completely first. If the output cannot be completed at one time, please output it in multiple times. And output the tables, and the titles and content descriptions of the pictures.

Result

Kimi

  • The translation was completed very smoothly
  • The translation of professional terms is relatively good
  • It will output a normal paragraph in Markdown list format, which is not necessarily a good or bad thing, it depends on the person.
  • The typesetting of very complex tables in the appendix is ​​not good, but there is no problem with general tables.

Doubao

  • The interface is very friendly, with a dialog box on the left and the uploaded document displayed on the right.
  • The translation was completed very smoothly
  • The translation of professional terms is relatively good
  • The paragraph output maintains the original format and will not be changed to Markdown list format like Kimi
  • The output of complex tables is slightly worse than Kimi, and there is no problem with simple tables

Tongyi Qianwen

  • Regardless of whether the deep thinking mode is turned on or off, the document content cannot be accurately extracted, and the location to be translated cannot be accurately located.
  • The Qwen open source model is very well done, but the official website does a very poor job at least in terms of document reading and extraction, and the instruction following is also not good.

Wenxin Yiyan

  • Wenxin Yiyan explicitly has a call tool switch on the interface, and calls Python to read the document content during the thinking process, but it still cannot read the document content correctly after many attempts, and finally fails.
  • Because it could not read the document content, it began to perform the popular and serious nonsense when outputting the results.

DeepSeek

  • The first two chapters were translated smoothly, and the output was not bad.
  • But when I started translating the third chapter, it got stuck, prompting: Server is busy
  • Switched to the R1 model, still prompting: Server is busy
  • It seems that DeepSeek’s server resources are indeed insufficient, and may have been used to train R2.

Tencent Yuanbao

  • Turned on the T1 deep thinking mode, and the first five chapters were translated smoothly. The output likes to refine and summarize, and the original text is not followed very well; when asked to translate the following content, a large amount of repeated original content is displayed in the thinking chain, and then it gets stuck.
  • Turned off the T1 deep thinking mode, the first four chapters were basically normal, and from the fifth chapter, it began to add content on its own, and the instruction following became worse.

Summary

  • The current best performers in document-level translation are Kimi and Doubao
  • The underlying model is very important, but for a real-world task, there are too many factors that affect the results, and the other companies still have a gap in this regard.