HY-MT Loader

The HY-MT Loader is the least glamorous node in this pack, and honestly the one you'll reach for first. It does exactly one thing - it finds a local Hunyuan Translation model, loads it into VRAM, and hands you a handle for the translator. By itself it renders nothing, translates nothing, and shows you a single dropdown. But it's the difference between your first translation taking ten seconds and taking ten minutes.

This pack (starsFriday/ComfyUI-HY-MT1.5, a fork of Tencent's official HY-MT repo) wraps Tencent's open translation models for ComfyUI. The star is Hy-MT2-1.8B - a "pocket-sized" 1.8B model that covers 33 languages and runs on modest hardware - plus the older HY-MT1.5 checkpoints the pack was named after. The Loader is the model half of that story; the HY-MT Translator is the other half.

How it works

When you open the node, it scans two folders under your ComfyUI models directory, in order: models/Hy-MT2-1.8B/ first, then models/HY-MT1.5/<model_dir> for legacy checkpoints. Everything it finds becomes a dropdown entry. Run it and it loads the tokenizer plus model via HuggingFace transformers - bfloat16 if you have CUDA, plain float32 otherwise, with device_map="auto" so the model spreads across whatever VRAM and RAM you have.

Two things are worth knowing about that load. First, it's cached per folder path and stays resident - run the Loader once, and every translator wired to it reuses the same weights instead of reloading them on each translation. Second, the model is loaded with trust_remote_code=True, which means HuggingFace runs the model's own code on your machine. That's normal for this kind of model, but it's why you should download weights from Tencent's official repos (or a source you trust) rather than grabbing random uploads.

The inputs and outputs

• model - the only input. An enum dropdown listing every model folder it found. Pick the one you want and forget it.

The single output is model, typed HYMT_MODEL. Don't be thrown off by the custom socket type: it's not a tensor, it's a (tokenizer, model) pair, and it only plugs into one thing - the model input on the HY-MT Translator. That's the entire wiring: Loader output → Translator input.

There's a gotcha hiding in the dropdown: if it reads <missing HY-MT models>, the node found nothing. That's not a bug - it's the signal that your weights aren't where it's looking.

Installing it

ComfyUI Manager is the easy path: search ComfyUI-HY-MT and install. Or from a terminal:

cd ComfyUI/custom_nodes
git clone https://github.com/starsFriday/ComfyUI-HY-MT1.5

Then restart ComfyUI. The repo itself declares no dependencies in pyproject.toml - it leans on transformers and torch, which a working ComfyUI install already has. The README's one caveat: Hy-MT2's model card recommends transformers>=5.6.0, and older versions may choke on the model's implementation. If loading throws, that's the first thing to check.

The actual heavy lift is the model download. Grab tencent/Hy-MT2-1.8B from HuggingFace and drop the folder into models/Hy-MT2-1.8B/ (or a HY-MT1.5 checkpoint under models/HY-MT1.5/<name>/). A 1.8B bf16 model is a few GB - nothing like the 45GB Hunyuan Video wants.

Common issues

• Dropdown shows <missing HY-MT models>: the folders aren't there or the names don't match exactly. Hy-MT2-1.8B (with the dash) under models/, or a subfolder under models/HY-MT1.5/.
• It keeps the model resident in VRAM. That's the cache doing its job. If you're tight on VRAM for the generation itself, consider not wiring the Loader at all - the Translator will auto-load the first model it finds and you can free it by reloading the workflow.
• Loading fails on Hy-MT2: check your transformers version against the >=5.6.0 recommendation before blaming the pack.

One honest caveat: this is a small, lightly-maintained pack from a hobbyist author (the same starsFriday who ships ComfyUI LoRAs on HuggingFace), and the community footprint is thin - the underlying Tencent models got a brief shout in a May 2026 roundup and little else. It works, it's local, and it's free per call. Just keep your model source trusted, because trust_remote_code doesn't forgive careless downloads.