Mặc Hi is a web app that translates Chinese to Vietnamese, specifically for online stories and novels — running 100% in your browser, on your own machine's GPU. No servers, no accounts, no API costs, no usage limits — your text never leaves your device. Download the model once, install it as a PWA, and thereafter translate completely offline.

Two models, switchable directly in the app

- MoxhiMT-30 — ~30 million parameters, download ~74 MB: lightweight, prioritizing speed.

- HachimiMT-60 — ~60 million parameters, download ~113 MB: a larger option for comparative translation.

Both are self-trained for narrative style, doing real neural translation rather than word-by-word dictionary lookup.

Where does Mặc Hi stand?

Not aiming to replace QuickTrans/VietPhrase or LLMs, it lies between those two groups: compared to QT/VP, Vietnamese sentences are more smoothly natural without needing to build a rule repository — but QT/VP is still superior due to being ultra-lightweight/ultra-fast, particularly when absolute control over proper nouns and address forms is needed. Compared to general-purpose LLMs, Mặc Hi is hundreds of times smaller — while LLMs are more suitable for difficult segments that require deep contextual reasoning or style editing.

Runs on any GPU — not locked to CUDA

Unlike native machine translation tools that are tightly integrated with the CUDA/NVIDIA ecosystem, Mặc Hi runs on the open WebGPU standard, automatically recognizing the graphics backend of each machine (D3D12 on Windows, Metal on Apple, Vulkan on Linux/Android).

Actual speed

Speed depends on the GPU and the length of the text — the numbers below were measured on specific devices:

- Laptop RTX 5070 Ti: translates a novel of 2.23 million characters in ~17 seconds (~129,000 source characters/second, end-to-end).

- On the public benchmark (standard 21,383 sentences): RTX 5090 ~295,000 tokens/second; iPhone 17 Pro Max ~11,700 tokens/second. For more details, visit https://moxhi.vietphrase.app/board

- Integrated GPUs and mid-range phones: a few thousand characters/second — still plenty of speed for "translate as you read."

Runs on Chrome / Edge 113+ (Windows, macOS, Linux, Android) and Safari 26+ (iPhone, iPad, macOS).

The story behind

With small Marian translation models, the usual execution through ONNX Runtime Web and Transformers.js hasn't exploited the GPU very effectively. Mặc Hi chose a more specialized direction: the entire inference engine is rewritten in WebGPU/WGSL to optimize inference speed.

On the RTX 5070 Ti Laptop, Mặc Hi's custom decode loop is around 10–32 times faster than ONNX Runtime Web running WebGPU/fp16 in tested workloads.