Chinese AI Glossary (中英对照)

Why a Chinese-first, bilingual AI glossary

Most of the foundational vocabulary of artificial intelligence — tokens, embeddings, fine-tuning, retrieval-augmented generation, chain-of-thought — was coined in English. Yet a huge share of everyday AI use now happens in Chinese, on models such as DeepSeek, Qwen (通义千问), Doubao, Kimi, ERNIE and Zhipu, and in Chinese-language tutorials, product pages and news. That mismatch creates a recurring friction: you read 上下文窗口 in a Chinese article and 「context window」 in an English paper without realising they are the same thing, or you know what 微调 means but freeze when a model card says "fine-tuning". This glossary exists to close that gap. It is Chinese-first by design — every entry leads with the Chinese term and a plain-Chinese explanation — but it always pins the English cross-reference right next to it, so the two halves of the field reconnect.

The list covers more than forty of the terms that come up most often when you actually use a large language model: 大模型 and Token, 提示词 and 系统提示, 温度, Top-p and 最大输出长度, 思维链 and 少样本提示, RAG and 向量数据库, 智能体 and 工具调用, 幻觉, 对齐 and 护栏, 蒸馏, 量化, 嵌入, MoE and 推理模型. Each definition is deliberately short — one or two sentences that capture what the term means and why it matters, without formulas or unnecessary jargon. That makes the glossary equally useful to a beginner trying to follow an AI news story and to an engineer who simply needs to confirm the standard Chinese rendering of an English term before writing documentation.

How to use the glossary to read and prompt better

Using the glossary is straightforward. Open the page and the whole list is there, sorted A–Z by the English term so it lines up with English sources. To narrow it, type in the search box: it matches the Chinese term, the English term and the words inside each explanation, so 「微调」, "fine-tuning" and even a fragment of the definition all find the same card, and a live counter shows how many terms match. Because nothing is sorted away when you clear the box, you can also just browse the full list end to end as a quick primer on the vocabulary of modern AI.

Knowing the right words pays off the moment you start prompting. Once "context window" and 上下文窗口 are clearly the same limit, you stop pasting more text than a model can hold; once you understand 温度 and Top-p, you can tune a model's creativity on purpose instead of by trial and error; once 思维链 and 工具调用 are concrete, you can read a model's capabilities and design better prompts and agents. And because the entire tool runs locally in your browser — no network request, no model call, no stored search — you can keep it open as a reference tab and consult it as often as you like while you read, write or build, with complete privacy. Treat it as a companion to the Chinese models you already use: when a term in a tutorial, a release note or an API reference stops you, look it up here, see both languages at once, and carry on.