77Fox's Windows desktop app can run five translation models on your own PC: M2M100, Hy-MT2-1.8B, Hy-MT2-7B, Index-Translate-2B and Index-Translate-9B. Translation runs locally, the text stays on your computer, and no AI credits are used; pick a model by the languages you translate first, then by your GPU. The models run inside the Windows desktop app, and the browser extension can use them through it; the web console and the Android app can't.
Plenty of people would rather not send contracts, client files or internal documents to any cloud service. On a plane or an intranet-only network, cloud translation can't connect anyway, and pay-per-use APIs add up. The hard part is choosing: downloads differ almost fivefold, language coverage runs from 33 to 150, and GPU support depends on who made your graphics card.
The five models at a glance
| Model | Download | Languages | License | Publisher | GPU acceleration |
|---|---|---|---|---|---|
| M2M100 (418M) | About 1.8 GB | About 100 | MIT | Meta | NVIDIA only |
| Hy-MT2-1.8B | About 1.1 GB | 33 | Apache-2.0 | Tencent Hunyuan | NVIDIA, AMD, Intel |
| Hy-MT2-7B | About 4.3 GB | 33 | Apache-2.0 | Tencent Hunyuan | NVIDIA, AMD, Intel |
| Index-Translate-2B | About 1.2 GB | 150 | Apache-2.0 | bilibili's Index team | NVIDIA, AMD, Intel |
| Index-Translate-9B | About 5.4 GB | 150 | Apache-2.0 | bilibili's Index team | NVIDIA, AMD, Intel |
Download sizes are the figures 77Fox shows in Settings. M2M100 also needs a Python environment, covered in the next section.
In one line each:
- M2M100: about 100 languages; needs a Python environment and runs on the CPU only with AMD or Intel graphics.
- Hy-MT2-1.8B: the smallest download of the five, 33 languages.
- Hy-MT2-7B: the same 33 languages in a file about four times larger.
- Index-Translate-2B: 150 languages in a file only slightly larger than Hy-MT2-1.8B.
- Index-Translate-9B: the largest file of the five, also 150 languages.
They live in Settings → 77foxAI → Local AI, under Local models. Each downloads separately; to remove one, right-click its row and choose Uninstall and delete files. Downloading isn't loading: a model enters memory only when something uses it. A stopped download resumes where it left off, and Change folder moves the model folder to a bigger drive.
Two things to check before downloading
The runtime each model needs
Hy-MT2 and Index-Translate share one runtime, labeled "llama.cpp runtime" in the app, installed with the first of them you download. The build for NVIDIA cards is about 630 MB; on other PCs it's a few dozen MB.
M2M100 uses a separate Python environment: Python 3.10–3.13 must already be on your PC, and 77Fox then installs PyTorch and other dependencies, about 3.4 GB on the NVIDIA route or about 950 MB on the CPU route. Parts already installed are skipped, and an NVIDIA card with an old driver takes the CPU route too. Without Python, you'll see "Python 3.10–3.13 not found. Install Python first, then set up"; an unsupported version gets its own message.
Where downloads come from
Hy-MT2 and Index-Translate download from Hugging Face only, pinned to fixed versions and checked after download; the llama.cpp runtime comes from GitHub. M2M100 has a Download source menu that defaults to ModelScope, a model hub based in China, and falls back to Hugging Face if that download fails; you can also pick Hugging Face directly.
On networks where Hugging Face or GitHub is restricted, Hy-MT2 and Index-Translate may fail to install; M2M100 can download from ModelScope.
Choose by GPU
| Your PC | When 77Fox picks automatically | Suggestion |
|---|---|---|
| NVIDIA GPU | With a recent driver, all five run on the GPU via CUDA | Choose by language; the 7B and 9B files are large, so leave room in VRAM |
| AMD GPU | Hy-MT2 and Index-Translate can use the GPU via Vulkan; M2M100 runs on the CPU | Prefer Hy-MT2 or Index-Translate |
| Intel GPU | Same as AMD | Same as AMD |
| No usable GPU | All five run on the CPU | Start with Hy-MT2-1.8B or Index-Translate-2B: smaller download, lighter footprint |
Vulkan acceleration on AMD and Intel needs the Vulkan runtime present in Windows. Without it, the CPU is used.
Setting Translation hardware
Translation hardware appears in Settings → 77foxAI → Local AI, in the screenshot translation settings and in the browser extension's settings. They're all the same setting.
- By default, 77Fox picks automatically and uses a GPU when one is usable, as in the table above.
- You can pick CPU or GPU by hand. The switch waits for the current translation to finish and survives a restart. To go back to automatic, click Auto select, which appears once you've made a manual choice. When the current model can't use the GPU, the GPU option is greyed out.
- The row shows both the device you chose and the one actually running, such as "Running on GPU". Selection translation and screenshot translation also tag the elapsed time with CPU or GPU for that run.
- If the GPU fails, you get an error; there's no quiet fallback to the CPU. Lock GPU and then switch to a model your card can't accelerate, such as M2M100 on an AMD card, and loading fails; set Auto select or CPU instead.
These choices are stored on this PC only and don't sync with your account.
Choose by language
- If everything you translate falls within Hy-MT2's 33 languages, choose Hy-MT2; 1.8B is the lightest of the five.
- If your source text often includes less common languages, choose Index-Translate, which covers 150.
- M2M100 sits in between at about 100 languages. Its catch is the runtime: Python required, GPU acceleration on NVIDIA only.
- Within each family, both sizes cover the same languages and differ in file size and memory use.
The language counts describe what each model covers. Through 77Fox, local models translate into 33 common languages, so a model with wider coverage mainly reads more source languages. A count also doesn't mean every language comes out equally well; running a paragraph of your usual material through each candidate tells you more than any table.
Memory: one model at a time, and Keep loaded
Of the five, only one model sits in memory or VRAM at any moment. When you switch, 77Fox confirms the old one has fully exited before loading the next.
Without Keep loaded, a model is released after 15 idle minutes, and the next translation waits for it to load again. Keep loaded lets you pick one installed model: it loads right away and loads again whenever 77Fox starts. Choose None to release it immediately. With Keep loaded on, if another feature borrows a different model, the kept model comes back afterwards. You'll find it in Settings → 77foxAI → Local AI and in the screenshot translation settings, and the tray menu has a Keep local translation model loaded switch.
| Keep loaded | Memory or VRAM | Wait before translating | Good for |
|---|---|---|---|
| None (default) | Released after 15 idle minutes | The first run after release waits for loading | Occasional use, tight memory |
| A chosen model | Held all the time | No loading wait | Heavy daily use with memory to spare |
Try to use the same model for screenshots, web pages and prompts. Only one fits at a time, so mixing models means repeated unloading and loading.
Where local models work
| Where | Local models? | Notes |
|---|---|---|
| Page and selection translation | Yes | Through the 77Fox browser extension, installed via the desktop app, which calls the paired desktop app over a local connection |
| Screenshot translation | Yes | In the Windows app; see screenshot translation |
| Prompt translation | Yes | In the prompt library; see prompt translation |
| Web console and Android app | No | Local options show "Requires desktop"; the Android app loads the web console |
| The standalone 77Translate extension | No | Only the free Google channel and your own keys; see 77Translate |
For browser use, keep the 77Fox desktop app running and paired. If it isn't, the extension temporarily switches to the first source you can use, in this order: your own key, 77Fox Cloud (signed in with credits left on your plan), then the free Google channel (when it isn't rate limited). The page shows a one-time notice such as "77Fox isn't running, so the local model is temporarily using …". Meanwhile page text goes to that source, and 77Fox Cloud uses credits. It switches back once the desktop app returns. If none of the three is available, the extension says the desktop app is offline. Screenshot and prompt translation run inside the desktop app, so this doesn't apply to them.
Extension setup is covered in local translation models in the browser extension, the engine types in choosing a page translation engine, and screenshots in translating text from a screenshot.
FAQ
Is it fully offline?
Downloading a model and its runtime needs internet; translation after that runs on your PC. Screenshot and prompt translation with a local model work with the network unplugged. Page translation reaches the desktop app over a local connection, so it works offline too while the app is running and paired; otherwise the extension switches to another source for the time being.
Can I use it on a laptop without a dedicated GPU?
Yes. All five models run on the CPU. With only Intel or AMD integrated graphics, the automatic choice may still run Hy-MT2 and Index-Translate through Vulkan; the settings show which device is actually used. If memory is tight, start with Hy-MT2-1.8B or Index-Translate-2B.
Why doesn't it fall back to the CPU when the GPU fails?
So you always know what is actually running. A quiet fallback would still produce translations while hiding that the GPU wasn't working, and you wouldn't know whether to fix a driver or change a setting. After the error, set Translation hardware to CPU and try again; if you locked the GPU by hand, Auto select works too.
Can I use these models commercially?
Only the license names are stated here: M2M100 is MIT, and Hy-MT2 and Index-Translate are Apache-2.0. Whether that fits your use depends on each publisher's license text; this isn't legal advice.
Do local models cost anything?
No AI credits are used. Only 77Fox Cloud translation is billed in credits (pricing); a local model costs disk space, one download and your PC's own computing time.