Zoey - 混元翻译器 (HY-MT1.5/MT2.0)

HunyuanTranslatorNode is the heavyweight answer when a translator in your workflow isn't allowed to be an API call. It runs a Tencent Hunyuan MT (machine-translation) LLM locally on your GPU, downloads the weights for you on first use, and translates across 20+ languages - including the tricky Chinese variants most tools skip: Traditional Chinese, Cantonese, Tibetan, Uyghur, Mongolian, Kazakh. If you're building Chinese↔anything pipelines or handling subtitles and dialogue that need genuine fluency rather than Helsinki-model guesswork, this is the node the pack wants you to reach for when its lighter PureTranslator isn't enough.

It ships in comfyui-ZoeyTool alongside PureTranslator, and the pairing is the story: the "pure" node covers ZH/EN/JA/KO with tiny models and APIs; this one covers everything with a real generative translator, at the cost of disk space and VRAM.

How it works

You pick a 模型版本: HY-MT1.5 or HY-MT2.0, in 1.8B / 7B / 70B sizes (default HY-MT2.0-1.8B). On first use the node downloads the matching checkpoint from Hugging Face (tencent/HY-MT2.0-1.8B and friends) into ComfyUI/models/LLM/<version>/, then loads it through transformers with trust_remote_code. It's cached per model+device after that, so first run pays for downloads and model load, subsequent runs are fast. Pick your hardware honestly: 1.8B runs on modest cards, 7B wants a real GPU, and the 70B variant is a datacenter play - don't click it on a 12GB card and expect it to work.

The inputs that make it feel like a translation tool rather than a dumb model wrapper:

• 术语干预 - terminology lines like AI -> 人工智能 / GPU -> 图形处理器; it bakes those into the prompt so proper nouns and jargon stay consistent. Huge for translating prompt templates or product copy where you've fixed your own glossary.
• 上下文 - paste prior text and the translation is context-aware (pronouns, tone) instead of sentence-by-sentence.
• 保留格式标签 - when on, it preserves <sn>…</sn>-style format tags through the translation, for code/serialization text you don't want mangled.
• 最大新令牌数 - generation length cap (64–2048, default 512). Raise it for long prose.

Outputs are 翻译结果 plus 耗时(秒).

Installing it

Same pack install as everything here:

cd ComfyUI/custom_nodes
git clone https://github.com/liangzoey/comfyui-ZoeyTool.git
cd comfyui-ZoeyTool
pip install -r requirements.txt

Restart ComfyUI (or ComfyUI Manager → comfyui-ZoeyTool). Plan for a big first-run download - these are multi-gigabyte Hugging Face repos, and on slow connections the node's download step can look hung when it's really just moving several GB. It has a config.json alongside the node code if you want to point LOCAL_MODEL_ROOT at a different models folder than the default.

Where people get burned

VRAM is the obvious one - don't assume "1.8B default" means light; a 1.8B transformer still wants a few GB comfortably above torch's baseline. Second, Tencent's Hunyuan releases generally ride Tencent's community license, which historically excludes some territories (EU/UK/KR) - worth a glance at the model card before you build a commercial pipeline on it. And third, HF download failures mid-way are a real thing; the node retries with force_download, but if your network blocks Hugging Face entirely, you'll want to fetch the repo by hand into models/LLM/<version>/ and let it find it locally. There's essentially no community chatter about this node yet (it's new and niche), so when it misbehaves, the console logs - which are unusually detailed, down to the resolved model path - are your best friend.