A Chinese character is worth about 0.6 of a token, which is the same thing as saying a token is worth about 1.6 Chinese characters. Four Chinese model vendors publish that figure in their own developer documentation and they agree: DeepSeek writes that one Chinese character is about 0.6 tokens, Zhipu gives the ratio as about 1 to 1.6, MiniMax says 1,600 Chinese characters consume about 1,000 tokens, and Moonshot's Kimi puts one token at roughly 1.5 to 2 Chinese characters. So a page of 1,600 Chinese characters costs you about a thousand tokens, not sixteen hundred — and the figure is model-specific, which every one of those four pages says, each of them pointing you at the usage figure the API itself returns.
That matters because the number you will find repeated on the open web is 1 character = 1 token, and it is wrong by about sixty per cent in the direction that costs you money. Several of the pages that print it also print "1.5 to 2 tokens per character" further down, which is wrong in the other direction. If you want the mechanism rather than the number, our guide to what tokens and context actually are covers it, and our tokenisation visualiser will split a sentence in front of you.
The rest of this guide is about the words. We collected the published developer documentation of eight Chinese model vendors — DeepSeek, Alibaba's Qwen, Moonshot's Kimi, Zhipu GLM, ByteDance's Doubao, Baidu's ERNIE, Tencent's Hunyuan and MiniMax — 32 documents, 730,128 Chinese characters, and counted which Chinese word each of them actually writes. What follows is what the vendors write, not what a dictionary says they should.
The word for token is "token"
Start with the word you will meet first, because it is the strangest result in the set. All eight vendors write token in Latin letters in their running Chinese prose. Not one of them writes 词元, which is the term China's national approved-terminology service actually lists for it. Across all 730,128 characters we collected, 词元 appears zero times.
Meanwhile 令牌, which is the other approved rendering and the one a bilingual glossary is most likely to hand you, appears 27 times — and when we classified each occurrence by the words around it, 18 of the 27 mean an access credential: 访问令牌, API 令牌, 刷新令牌. Two are the model sense and the remaining seven are ambiguous. So if you are reading a Chinese vendor's docs and you hit 令牌, the safe first guess is that somebody is talking about authentication, not billing.
This is an easy trap to fall into; we did. Our own Chinese AI glossary listed 词元 as the headword for token, which is defensible as terminology and misleading as advice; it now leads with the word the documentation uses and keeps 词元 as the approved form. This pattern holds across the subject: there is an official Chinese vocabulary for this field, and then there is the vocabulary the field actually uses. For several key terms, they don't match.
One small pleasure from the corpus: DeepSeek and Zhipu define token in an identical sentence, word for word — "Token 是模型用来表示自然语言文本的基本单位…" and onward. That sentence appears in exactly two of the eight documentation sets and nowhere else. The standard Chinese definition of a token is, in part, a shared text.
If you learn one word, learn 大模型
The textbook translation of "large language model" is 大语言模型, and it is not what anybody writes. In the documentation we collected, 大模型 — literally "big model", with the "language" dropped — outnumbers it by roughly 40 to 1, and is used by seven of the eight vendors. 大型语言模型, the Taiwan-standard form, appears six times in total.
A few more where one form clearly leads:
| English | What the vendors write | The runner-up |
|---|---|---|
| large language model | 大模型 — 7 of 8 vendors, 1,000 occurrences | 大语言模型, 25 |
| agent | 智能体 — 7 of 8, 753 | 代理, 53 |
| chain of thought | 思维链 — 7 of 8, 224 | nothing else in use |
| multimodal | 多模态 — 7 of 8, 359 | 全模态, 37 |
| prompt | 提示词 — 6 of 8, 230 | 提示语, 6 |
| retrieval-augmented generation | 知识库 — 6 of 8, 414 | 检索增强生成, 8 |
| computing power | 算力 — 5 of 8, 107 | 计算能力, 0 |
Two of those rows stand out. Vendors overwhelmingly say 知识库, "knowledge base", where an English document would say RAG — they are describing the product feature rather than the technique, and the technical term barely appears. And 算力, the everyday word for compute, is not an approved term at all: the national list records it as the short form of 计算能力, which appears zero times in the corpus.
Context length, not context window
The clearest point of agreement among vendors is on context length. 上下文长度 is the only Chinese form that all eight vendors use. 上下文窗口 is used by four of them. Baidu, ByteDance, DeepSeek and Tencent use 上下文长度 exclusively.
So if you are writing for a Chinese audience, or reading a spec sheet, 上下文长度 is the safe form — and if you are comparing numbers under that heading across vendors, be careful, because the same phrase gets attached to several different limits. Our guide to the four different length limits one model has is about exactly that confusion, and it does not go away in translation.
Where the vendors genuinely disagree
On fine-tuning — adapting a trained model to your own data — vendors disagree sharply. Baidu writes 精调 and never 微调, with 92 occurrences of the first and none of the second in the collected corpus (and on a single content page of its model list, 106 against zero). Zhipu does the opposite, writing 微调 101 times and 精调 once. 调优 is used by six vendors. There is no house form here; there are three, and which one you see tells you whose documentation you are reading.
Mixture of experts is mostly left as MoE. When it is translated, it splits 混合专家 against 专家混合 — and Kimi uses both, for the same concept, on different pages of its own documentation. Reasoning models split three ways too: 推理模型 (six vendors), 深度思考模型 (five), 思考模型 (four).
Six English words are left untranslated by a majority of vendors in running Chinese prose: token, prompt, agent, embedding, MoE and MCP. API parameter names like temperature and max_tokens are also in Latin letters, but those are code rather than a translation choice and should not be counted as one.
A limitation of our method: Baidu's and Tencent's documentation sites serve the same navigation tree with every page, so part of those two vendors' text is section headings rather than prose. Every Baidu figure above was re-checked against a single real content page in isolation and the direction of each claim survived.
Taiwan writes it differently, and sometimes not at all
Taiwan's official terminology service is run by the National Academy for Educational Research, and it is genuinely official rather than a crowd-sourced list. The everyday differences are the ones you would expect: 人工智慧 rather than 人工智能 (unanimous across 86 entries), 演算法 for algorithm, 軟體 for software, 資料庫 for database. Then it gets more interesting.
- 智能體 returns nothing at all in Taiwan's database. The approved forms are 智慧型代理, 人工智慧代理, 智慧主體 and 智慧代理人. This is a wider gap than 智能/智慧, because the mainland word simply does not exist on the Taiwan side.
- 大模型 has no AI sense in Taiwan's database either — the only hits are for an aerospace mould. Nor does 算力, whose only hits are in computational mechanics.
- Taiwan's approved form for mixture of experts is 專家混合, with the word order reversed from the commoner mainland 混合专家; attention mechanism is recorded as 注意機制 as well as 注意力機制.
- Token has five approved Taiwan renderings — 符記, 訊標, 權杖, 詞元, 代幣 — which is another way of saying the word has not settled.
One thing that is not a clean split, despite being widely presented as one: 提示語 against 提示詞. Taiwan's own entries carry both, and on the mainland side 提示语 is used by two of the eight vendors. Both forms exist on both sides, and 提示詞/提示词 leads on both. If you have read that 提示語 is "the Taiwan word", that is tidier than the evidence.
The official lists have not caught up, on either side
We probed both national services directly, with control terms included so that a broken query could not be mistaken for a missing term — a first attempt failed in exactly that way and was thrown out.
On the mainland list, of 35 terms from the large-model era, 17 are present and 18 are absent. Absent: 大语言模型, 大模型, 生成式人工智能, 提示词, 提示工程, 预训练, 上下文窗口, 思维链, 检索增强生成, 分词, 向量数据库, 算力, 混合专家, 扩散模型, 工具调用, 越狱, 知识蒸馏 and 提示词工程. Present: 微调, 幻觉, 智能体, 上下文, 多模态, 词元, 令牌, 嵌入 and others.
On the Taiwan side, of 32 English search strings from the same era, eleven returned nothing, and after folding spelling variants seven distinct concepts have no Taiwan entry whatever: retrieval-augmented generation, context window or context length, vector database, RLHF, knowledge distillation, semantic search and jailbreak. A detail that says something about how new all this is: the Taiwan discipline glossary carrying the most AI terms is the one for civil engineering surveying and graphics, followed by visual arts — not a computing glossary at all.
The practical rule is this: for terms older than about 2020 (like 注意力机制, 嵌入, or 微调), the approved and used forms generally agree, and either will be understood. For anything the large-model era brought with it, write what the vendors write, and expect it to move.
The character-per-token figures are quoted from each vendor's own developer documentation: DeepSeek's token-usage page, Zhipu's and Moonshot's billing-concepts pages and MiniMax's billing notes. The word counts are ours, from 32 documents totalling 730,128 Chinese characters collected on 24 September 2026, counted longest-form-first with each match masked so that 上下文窗口 is never also counted as 上下文, and with byte-identical documents from one vendor counted once. The Taiwan forms come from the National Academy for Educational Research's terminology service and the mainland ones from the national approved-terminology service, both queried directly on the same date. All of this is a snapshot of documentation that changes often; treat the counts as of that date rather than as a standing fact.
View author profile → · Editorial policy
