Hangzhou, open weights, then a reasoning shock
DeepSeek grew out of High-Flyer, a Hangzhou quant fund, and published a run of open-weight models that the rest of the field had to take seriously. DeepSeek-V2 and V3 showed that a well-trained mixture-of-experts model could rival much larger Western labs. DeepSeek R1, in 2025, made chain-of-thought reasoning a public event: a model that showed its work, at a price that forced every other house to answer. The company stayed stubbornly technical. The language it knows in its bones is 普通话.
V3 for talk, R1 for work you can see
DeepSeek V3 is the chat seat. DeepSeek R1 is the reasoning seat people still name when they want visible work. On this rail the live job is speech, not a proof contest and not a research crawl. Proof-shaped questions stay with house math. A sourced 中文 dossier is Kimi.
What Mandarin owns
Mandarin is a live chat partner in 普通话. It writes characters. It does not extract a government form — that is GLM Form. It does not mix English comments into a repo — that is Qwen Code. It does not run a WeChat workbench — that is Hunyuan.
How to ask in 普通话
@mandarin, or pin the desk from Floor. A localization of a help center is Translate. If you want sources from the 中文 web, leave this rail and open Kimi.