Soft Translation

把鼠标放到任意段落上，按 Shift，译文以灰色斜体出现在原文正下方—— 原文一字不动。 为什么安装它 · 永久免费：没有订阅、没有按词计费、没有广告版。用它自己的本地模型 （llama.cpp / Ollama），你有一台能跑模型的机器，它就是免费的。 · 隐私握在自己手里：你翻译的文本只发往你自己填写的服务器地址， 不经过任何第三方翻译 API，开发者也接触不到你读到的任何内容。 · 不打扰：原文页面绝不被修改，译文柔和地插入在段落下方， 并镜像原文的对齐与缩进，原文译文可以随意来回看。 · 对学术内容友好：自动识别 Markdown 与数学公式，LaTeX 公式渲染为 MathML、原样保留，绝不被翻译或破坏。 用途 · 原地阅读外文网页：arXiv 论文、技术文档、博客、新闻…… 不用复制文字、不用切换应用。 · 一键完成：鼠标悬停在段落上按 Shift，译文即刻出现； 翻译失败时显示红色提示，5 秒后自动消失。 · 带上下文翻译：自动附带周边段落作为消歧上下文，供模型参考， 可开关、长度可调。 灵活配置 · 支持 llama.cpp（/v1/chat/completions）与 Ollama（/api/generate）， 任何 OpenAI 兼容地址均可。 · "测试连接"一键拉取模型列表；目标语言、提示词模板、上下文均可配置。 · 端点可以是本机 127.0.0.1、局域网共享机器，或任何远程服务器。 前提：一台运行中的 OpenAI 兼容服务（llama.cpp / Ollama）及已加载的模型。 ──────────────────────────── Soft Translation: hover over any paragraph, press Shift, and the translation appears right below the original — in muted italic style, with the original text untouched. Why install it - Free forever. No subscription, no per-word billing, no ads. It runs on your own local model (llama.cpp or Ollama), so the only cost is the hardware you already have. - Your privacy, your endpoint. The text you translate goes only to the server you configure yourself. It never passes through a third-party translation API, and the developer never sees what you read. - Non-intrusive by design. The page is never modified; the translation is softly inserted below each paragraph, mirroring its alignment and indentation. - Built for technical content. Markdown and math are detected automatically: LaTeX formulas are rendered as MathML and kept exactly as they are — never translated or mangled. What it does - Read foreign-language pages in place — arXiv papers, documentation, blogs, news — without copying text into another app. - One keystroke: hover over a paragraph, press Shift, the translation appears instantly. Failures show a brief red notice that auto-dismisses. - Context-aware: surrounding text is automatically provided to the model for disambiguation (toggleable, size adjustable). Flexible setup - Works with llama.cpp, Ollama, or any OpenAI-compatible endpoint. - One-click "Test connection" loads the model list; target language, prompt template, and context are all configurable. - Point it at 127.0.0.1, a machine on your LAN, or any remote server. Requires: a running OpenAI-compatible server (llama.cpp / Ollama) with a model loaded.