The popup opens from the extension's icon in the browser toolbar — if it isn't there, click the puzzle-piece icon and pin Dual Subtitles.

弹窗从浏览器工具栏里扩展的图标打开——没看到图标就点拼图图标，把 Dual Subtitles 钉上去。

1. Open the extension popup → set Engine to Your own key (LLM / DeepL) → click Configure….
2. 打开扩展弹窗 → 把「翻译引擎」切到「自带 Key（LLM / DeepL）」 → 点「配置…」。
3. Pick a provider in the list on the left — fourteen of them, so the one this page starts with, Model Studio, is eighth; scroll. Paste your API key, leave the model on the recommended one. Want a different one? Hit List models with my key (List available models on a provider that needs none) and choose from what your key can actually reach — the list comes from the provider, so it says what that key reaches right now.
4. 在左边列表里选一个服务商——一共十四家，这一页最先讲的「百炼」排在第八，往下翻。粘贴你的 API Key，模型保持推荐值即可。想换别的？点「用我的 Key 拉取模型」（免 Key 的服务商上叫「拉取可用模型」），从你这把 Key 真正能用的模型里挑——列表是服务商当场给的，反映这把 Key 现在够得着什么。
5. Click Save and test. Chrome asks once for permission to reach that provider's domain — allow it, and you should see “Connected”.
6. 点「保存并测通」。Chrome 会问一次是否允许访问该服务商的域名——同意后应该看到「连接成功」。

Best pick inside mainland China, and the fastest of the three in our measurements. One key gives you both the Qwen models and DeepSeek models hosted there.

中国大陆首选，也是我们实测里最快的一家。一把 Key 同时能用平台上的通义千问系列和 DeepSeek 系列模型。

1. Sign in to the Alibaba Cloud Model Studio (Bailian) console with your Alibaba Cloud account, and activate the model service if it asks you to.
2. 用阿里云账号登录百炼控制台，如果提示需要开通模型服务，先开通。
3. In the console sidebar open System tools → API-KEY and create a key. (The link above goes straight there; console layouts do get reshuffled, so trust the link over the menu names.)
4. 在控制台左侧栏打开「系统工具 → API-KEY」，创建一把 Key。（上面的链接直达该页；控制台偶尔改版，菜单名不对就以链接为准。）
5. Copy the key — it starts with sk-. Paste it into the extension and press Save and test.
6. 复制这把 Key（以 sk- 开头），粘贴进扩展，点「保存并测通」。

• Recommended model: qwen-flash — the default, and by far the quickest we measured. deepseek-v4-flash also works well if you prefer DeepSeek's wording.
• 推荐模型：qwen-flash——默认值，也是我们实测最快的。想要 DeepSeek 的语感，就在拉取的列表里选 deepseek-v4-flash（没拉列表也可以直接手填），一样能用。
• Registered outside mainland China? Model Studio runs a separate international platform, and a key from one side is refused by the other. Pick International in the Site row before you press Save and test — the console and pricing links follow your choice.
• 账号不在中国大陆注册的：百炼另有一套独立的国际站，两边的 Key 不通用。按「保存并测通」之前先在「站点」那一行选「国际」——控制台和资费链接会跟着你的选择走。
• Keep the default endpoint. A workspace-scoped key (one that starts sk-ws-) works fine on the standard endpoint, so you do not need the workspace-specific host — and the standard one proved more reliable in testing.
• 接口地址保持默认。工作空间级的 Key（以 sk-ws- 开头）在标准接口上照样能用，不需要填工作空间专属域名——而且实测标准接口更稳。
• Avoid “thinking” models for subtitles: they spend most of their time reasoning. The extension already turns that off for this provider where the API allows it.
• 字幕别用推理（thinking）类模型，它们大部分时间花在思考上。这一家的接口支持关闭思考，扩展已经自动帮你关了。

An easy key to obtain — a couple of clicks in Google AI Studio, no separate billing setup to get started.

Key 很容易拿——在 Google AI Studio 点两下就有，开始用不需要另外配置付费。

1. Open Google AI Studio and go to the API Keys page.
2. 打开 Google AI Studio，进入 API Keys 页面。
3. Click Create API key and copy the value.
4. 点 Create API key，复制生成的值。
5. Paste it into the extension and press Save and test.
6. 粘贴进扩展，点「保存并测通」。

• Recommended model: gemini-flash-latest — the default. Do not pin a specific version: we found gemini-2.0-flash already answering “quota exceeded” on free keys, which looks exactly like a broken extension.
• 推荐模型：gemini-flash-latest——默认值。别把版本号钉死：我们实测 gemini-2.0-flash 在免费 Key 上已经直接返回「额度用尽」，看起来就像插件坏了。
• Not available in every country. If the test fails with a network error while other sites work, this is the likely reason.
• 并非所有国家/地区都能访问。如果别的网站都正常、只有测通报网络错误，大概就是这个原因。

Not a chat model — a dedicated translator. Wording is conservative and consistent, which many people prefer for subtitles. It also gets the per-line alignment right, because its API accepts the surrounding sentence as context.

它不是聊天模型，而是专业翻译引擎。用词稳、前后一致，很多人就喜欢这种字幕。它也能做到逐行对齐，因为它的接口支持把整句作为上下文传进去。

1. Sign up for a DeepL API plan — the Free one is enough. Note this is the API product, not the website translator or the desktop app.
2. 注册 DeepL API 方案，Free 版就够。注意要的是 API 产品，不是网页翻译或桌面客户端。
3. In your DeepL account page, open the section for API keys and copy the authentication key.
4. 在 DeepL 账户页面里找到 API 密钥那一节，复制认证密钥。
5. Paste it into the extension and press Save and test. There is no model to choose for DeepL.
6. 粘贴进扩展，点「保存并测通」。DeepL 不需要选模型。

• The single most common mistake: a DeepL Pro subscription (website / apps) is a different product from a DeepL API account, and the key you see in a Pro account does not work here. A 401/403 almost always means this.
• 最常见的错就这一条：DeepL Pro（网页版／客户端订阅）和 DeepL API 是两种不同的产品，Pro 账户里看到的密钥在这里用不了。报 401/403 基本都是这个原因。
• A key ending in :fx is a Free-plan key. The extension reads that suffix and talks to the right endpoint automatically — you never pick one.
• 以 :fx 结尾的是 Free 版 Key。扩展会读这个后缀自动选对应的接口地址——你不用管。
• Signing up needs a card. The API Free plan still asks for a VISA/Mastercard issued in a region DeepL supports — cards issued in mainland China are generally not accepted, which is why many users here end up on an LLM instead. DeepL's own pricing page has the current allowance.
• 注册要绑卡。API Free 方案仍然需要一张 DeepL 支持地区发行的 VISA／Mastercard——中国大陆发行的卡普遍不被接受，所以国内用户往往最后选大模型那条路。额度以 DeepL 自己的定价页为准。
• DeepL covers every language this extension offers. If that ever changes, you get a clear message asking you to change the language or the engine, and nothing is sent.
• DeepL 覆盖本扩展提供的全部译文语言。万一哪天不再覆盖，扩展会直接提示你换语言或换引擎，请求不会发出去。

Every endpoint below was probed and answered — the host and path are right. What we could not check without a key is the model name and whether the provider lets a browser extension call it directly. So don't guess a model: press List models with my key (List available models on a provider that needs none) in the settings page and pick from what your own account returns.

下面每一家的接口我们都探过、都有响应——域名和路径是对的。没有 Key，我们验证不了的是两件事：模型名，以及服务商允不允许浏览器扩展直连。所以别猜模型名：在设置页点「用我的 Key 拉取模型」（免 Key 的服务商上叫「拉取可用模型」），从你自己账号返回的列表里选。

A model running on your own computer costs nothing per video and sends nothing anywhere. The trade is speed and quality: a 7B model (seven billion parameters — the usual way of saying how big a model is) on a reasonably recent laptop usually translates a line in under a second once it is warm, and reads about as well as a cheap cloud model — how fast it really is depends on your computer. Anything smaller than 7B tends to mix languages mid-sentence, which is worse than not translating at all.

跑在自己电脑上的模型，看多少视频都不花钱，字幕也不会离开这台机器。代价是速度和质量：一个 7B 模型（70 亿参数，「7B」是说模型多大的惯用写法）在还算新的笔记本上热起来之后，通常一句话不到一秒，读起来大致相当于一个便宜的云端模型——具体多快，看你的电脑。比 7B 更小的，常常一句话里中英文混着出，那还不如不翻。

The model does not run by itself: you need a program that downloads models and runs them. Two free ones are easy to set up — Ollama, a small command-line tool, and LM Studio, the same idea with a window. Whichever you pick, the one thing that catches everybody is the same, and it is in step 2.

模型自己不会跑，要有一个负责下载和运行它的程序。免费又不难装的有两个：Ollama（一个命令行小工具）和 LM Studio（同样的东西，带窗口）。不管你选哪个，卡住所有人的都是同一件事，就在第 2 步。

Ollama

Ollama

1. Install it, then choose a model and pull it. Download Ollama from ollama.com and install it. It comes with no model of its own: models are pulled separately, a few gigabytes each, and the part after the colon in a name is the size class — 7b means seven billion parameters. Three things decide which one to pull:

1. 装上它，再挑一个模型拉下来。去 ollama.com 下载安装。它本身不带模型：模型要另外拉，一个几 GB；名字里冒号后面是大小档位，7b 就是 70 亿参数。拉哪个，看三件事：

• Size. 7B is the floor for subtitles — smaller ones mix languages, as above. Bigger reads better but is slower and needs more memory: count on the machine having a few gigabytes more memory than the model file, so a 4.7 GB model wants 8 GB and an 8 GB model wants 16 GB.
• 大小。字幕翻译 7B 起步——更小的会混语言，上面说过。更大的译得更好，但更慢、更吃内存：内存要比模型文件多出几 GB 才够，4.7 GB 的模型要 8 GB 内存，8 GB 的模型要 16 GB。
• Your languages. Every model's page on ollama.com lists what it speaks.
• 会不会你的语言。ollama.com 上每个模型的页面都写了它支持哪些语言。
• Not a thinking model. Names like qwen3 or deepseek-r1 think before they answer, and a subtitle line cannot wait for that — the same reason as the cloud note above.
• 别选会先思考的。名字里带 qwen3、deepseek-r1 这类的会先想再答，一句字幕等不起——和上面云端那条是同一个理由。

Then, in a terminal (the sizes and language counts above are ollama.com's figures at the time of writing — the model's own page is the current word):

然后在终端里跑（上面的大小和语言数是写这页时 ollama.com 上的数字，以模型自己的页面为准）：

ollama pull qwen2.5:7b

ollama list shows what is installed. Switching later is just pulling another one and picking it in step 3.

装了哪些，ollama list 会列出来。以后换模型，就是再拉一个，然后在第 3 步里选它。

2. Start it so the browser is allowed to talk to it. This is the step everybody misses. Ollama refuses requests from a browser extension unless you tell it not to, and the failure looks like the extension is broken rather than like a permission problem. Quit Ollama if it is already running, then start it like this (Mac and Linux — Windows is the next paragraph):

2. 启动它，并且允许浏览器连它。这一步是所有人都会漏掉的。默认情况下 Ollama 会拒绝来自浏览器扩展的请求，而且失败的样子看着像是扩展坏了，不像是权限问题。如果它已经在跑，先退出，然后这样启动（Mac 和 Linux；Windows 看下一段）：

OLLAMA_ORIGINS="chrome-extension://*" ollama serve

On Windows, set OLLAMA_ORIGINS to chrome-extension://* in the system environment variables and restart Ollama.

Windows 上把系统环境变量 OLLAMA_ORIGINS 设成 chrome-extension://*，然后重启 Ollama。

On a Mac that line lasts only until the terminal closes. To have the menu-bar app allow the extension every time it starts, run this once, then quit and reopen Ollama — it is how Ollama's own FAQ sets its variables:

Mac 上那行命令只管这一次，终端一关就没了。想让菜单栏里的 Ollama 每次启动都放行扩展，跑一次下面这行，再退出重开 Ollama——这是 Ollama 官方 FAQ 设环境变量的做法：

launchctl setenv OLLAMA_ORIGINS "chrome-extension://*"

3. Point the extension at it. In the settings page's Translation service pane — the page that Configure… in the popup opens — pick Ollama (local) from the provider list. There is nothing to type: this entry is wired to http://localhost:11434/v1, Ollama's own port, and it has no key field because Ollama does not use one. If yours listens somewhere else, use Custom (OpenAI-compatible) at the bottom of the list instead — that one has an address field.

3. 把扩展指过去。在设置页的「翻译服务」面里（就是弹窗里点「配置…」打开的那一页），从服务商清单里选 Ollama（本机）。没有东西要填：这一项直接连 http://localhost:11434/v1，也就是 Ollama 自己的端口；它也没有 Key 那一栏，Ollama 本来就不用 Key。如果你的 Ollama 换了端口，就用清单最下面那一项自定义（OpenAI 兼容）——那一项有地址栏。

Press List models with my key (List available models on a provider that needs none) and the models you have pulled appear — the list comes from your own machine, so it says what is actually installed. Pick one, then press Save and test. A green line means it answered.

点用我的 Key 拉取模型（免 Key 的服务商上叫「拉取可用模型」），你已经拉下来的模型就会列出来——这份清单来自你自己的机器，所以它说的就是你真正装了什么。选一个，然后点保存并测通。出现绿色那行就是通了。

If step 2 was skipped, this is where it shows: the test says it cannot reach the address, and names the two reasons — the server is not running, or it has not been told to allow the extension.

要是第 2 步漏了，就会在这里现形：测通会说连不上这个地址，并且把两个原因都说出来——服务没在跑，或者它还没允许扩展访问。

Running it somewhere else? The preset's address is fixed on purpose. If your Ollama listens on another port, or you reach a model server over a tunnel, use Custom (OpenAI-compatible) instead and type the address there — everything else on this page applies unchanged.

跑在别的地方？这个预设的地址是固定的。如果你的 Ollama 换了端口，或者你要通过隧道连一台别的机器，就改用「自定义（OpenAI 兼容）」，把地址填进去——这一页其余的内容照旧适用。

LM Studio

LM Studio

Same three steps, different switch. Download a model in LM Studio's own interface (the advice on size and languages above applies unchanged), open the Local Server tab, and turn on CORS before you start the server — that switch is this program's version of the line above, and without it the browser is refused in exactly the same way. In the extension, pick Custom (OpenAI-compatible) from the provider list — not Ollama — and give it the address http://localhost:1234/v1; the key stays empty.

同样三步，只是开关换了个地方。在 LM Studio 自己的界面里下载模型（上面关于大小和语言的建议照样适用），打开 Local Server 页，启动服务器之前先把 CORS 打开——这个开关就是上面那行命令的等价物，不开的话浏览器会以一模一样的方式被拒。在扩展里，从服务商清单里选「自定义（OpenAI 兼容）」——不是 Ollama——地址填 http://localhost:1234/v1，Key 同样留空。

Any other OpenAI-compatible server

其他任何 OpenAI 兼容的服务

That is what Custom (OpenAI-compatible) is for: a model server on another port, a machine on your network, a tunnel to your desktop, or a provider this page does not list. Four things decide whether it works.

「自定义（OpenAI 兼容）」就是干这个的：换了端口的本地服务、局域网里的另一台机器、连回家的隧道，或者这一页没列的某家服务。能不能通，取决于四件事。

When it does not work

不通的时候

The same local model can also write the summaries — the summary panel accepts whatever the Translation service pane is set to, including this.

同一个本地模型也可以拿来写总结——总结面板用的就是「翻译服务」面里配的那家，本地模型也算。

Read-aloud starts on your browser's own voices: nothing to sign up for, no key, and with the voices installed on your machine nothing leaves it (the browser's online voices, the ones named Google, send the line to Google to be spoken). Turn it on in the popup and it speaks each translated line as it appears. If you want better voices than your operating system ships, the Read aloud pane of the settings page (the gear in the popup's header, then Read aloud) takes a key from one of the providers below — it is a separate key from the translation one, even when the company is the same.

朗读默认用你浏览器自带的音色：不用注册、不用 Key；用本机装好的音色时什么都不离开这台电脑（浏览器的在线音色——名字带 Google 的那几个——会把这句发给 Google 来念）。在弹窗里打开开关，每句译文出现时就会念出来。想要比系统音色更好的，就在设置页的「朗读」面板里（弹窗页眉上的齿轮，再点「朗读」）填一把下面某家的 Key——它和翻译那把 Key 是分开的，哪怕是同一家公司。

Tick Read the translation aloud in the video and two sliders appear under it: Read-aloud volume, and Original audio — how far the video's own sound drops under the voice. The settings page has room for that one's full name, Original audio while speaking.

勾上「在视频里朗读译文」，下面才会出现两个滑杆：「朗读音量」，和「原声音量」——朗读时视频原声压到多低。设置页上写得下它的全名：「朗读时的原声音量」。

The popup has no preview: open the settings page's Read aloud pane (the gear in the popup's header), where Preview next to the voice speaks a sentence in your translation language before you commit to anything.

弹窗里没有试听：从弹窗页眉的齿轮进设置页的「朗读」面，音色旁边的「试听」会用你当前的译文语言念一句话给你听，不用先开着看视频才知道好不好。

If a provider's docs list more than one speech model, the same pane shows a Model row, and you can type a name yourself. When they ship a new model, type its name in and you can use it; you don't wait for an extension update.

服务商官方若写了不止一个语音模型，同一面板里会多一行「模型」，名字也可以自己填。他们新出了模型，把名字填进去就可以用，不用等扩展更新。

Click the small arrow on the extension's own subtitle button in the player's control bar (the one that is blue while the overlay is on — not YouTube's CC button) to open its menu, and pick Summary. The whole subtitle track becomes a one-line TL;DR plus chapters with timestamps, written in your translation language; click a timestamp to jump there. You need a service of yours: any own-key provider in the Translation service pane, or a local model — Ollama from the provider list, or LM Studio through the custom endpoint. The free engines only translate. They don't summarize.

点播放器控制栏里扩展自己那颗字幕按钮（开着时是蓝色的那颗，不是 YouTube 的 CC）上的小箭头打开菜单，选「总结」。整条字幕轨会变成一句摘要，加上带时间戳的章节，用你的译文语言来写；点时间戳就跳到那段。这需要你自己的服务：翻译面板里自带 Key 的任何一家，或本机模型——列表里选「Ollama（本机）」，LM Studio 走自定义接口。免费引擎只翻译，不总结。

With no service of your own set up, the panel opens on that instead: it says summaries need a service of yours and offers Set up a translation service, and there is no Summarize button until there is one.

还没配自己的服务时，面板开出来就是这句：它会说总结需要你自己的翻译服务，并给一颗「去配置翻译服务」；在那之前，「开始总结」这颗按钮根本不会出现。

Before you press Summarize, the panel names the provider and how many parts the track is split into. Until you press it, the track stays in the browser; you can cancel mid-way. Long videos are summarized in parts and then merged. If a part fails, it is written into the result rather than papered over.

按「开始总结」之前，面板会先写清用哪家、整轨分几段。你不按这个按钮，字幕不会离开浏览器；跑一半也可以随时取消。长视频会分段总结再拼起来，哪一段失败了就照实写在结果里，不会糊弄过去。

The extension tries to name the actual problem instead of just failing. Here is what each message means.

扩展会尽量说出到底哪里出了问题，而不是只报个失败。下面是每种提示的含义。

Nothing appeared at all? That is the other kind, and this table will not have it: no message means nothing got as far as failing. Open the settings page's About pane and work through the three checks under Something not working. The table below is for the other case — the key is set up and the extension is telling you what came back.

字幕根本没出来？那是另一回事，这张表里查不到：没有提示，说明它还没走到失败那一步。去设置页的「关于」面，照「遇到问题」下面那三步过一遍。下面这张表管的是另一种——Key 已经配好了，扩展在告诉你对面回了什么。

Still stuck? Copy the diagnostic info first (settings → About → Copy diagnostic info: version, engine state, what the current tab is doing — no key, no subtitle text) and paste it into your message; it saves a round of questions.

还是不行？先复制诊断信息（设置页 → 关于 → 「复制诊断信息」：版本、引擎状态、当前标签页在做什么，不含 Key 和字幕正文），贴进你的消息里，能省一轮来回。