divingclone/Hy-MT2-1.8B-NVFP4-vLLM

Hy-MT2-1.8B NVFP4 for vLLM

A calibrated 4-bit checkpoint of Tencent Hy-MT2-1.8B for the Hy-MT2 Windows project. Weights, configuration, tokenizer and chat template are stored directly at this repository root. No ZIP extraction is needed. Downloads are public and require no login.

Two model repositories are available:

• NVFP4 W4A4: RTX 50 fast profile, CUTLASS, 1,373,024,141 checkpoint bytes.
• GPTQ INT4 W4A16: symmetric group size 128, RTX 30/40 compatibility profile, Marlin; also usable on RTX 50, 1,300,648,141 checkpoint bytes.

Both retain BF16 tied embedding/output-head weights. The project's default KV cache is INT8 per token/head, a separate runtime setting. These are compressed-tensors safetensors checkpoints, not GGUF files.

Use

The Windows app downloads the selected repository's individual files from a pinned commit and checks each file's size and SHA-256. Choose one model for your GPU. CLI users run setup-model.cmd or setup-model.cmd --profile compat. Alternatively, use the official Hugging Face client:

from huggingface_hub import snapshot_download
model_path = snapshot_download("divingclone/Hy-MT2-1.8B-NVFP4-vLLM")
# Load model_path with the project's pinned native Windows vLLM runtime/plugin.

Validated runtime: community SystemPanic/vllm-windows 0.29.0, PyTorch 2.11.0+cu130, with the project's native Hunyuan model plugin. A standard repository layout does not guarantee compatibility with arbitrary upstream vLLM/Transformers versions. RTX 5090 was tested. RTX 30/40 support is based on runtime/kernel architecture support and still requires hardware validation. The current packaged runtime does not support GTX 10/16 or RTX 20.

Provenance and fidelity

Base revision: tencent/Hy-MT2-1.8B at 9a341cd1b679d3efd23b46e847b01745a71ed792. Calibration used 256 OPUS translation examples, disjoint from the 256-example evaluation set. File hashes are recorded in checksums.json. The INT4 repository also contains quantization.json with versions, calibration hashes and source/output weight hashes.

Using unquantized BF16 output as the teacher, the project measured chrF++ 84.97 for INT4/INT8 KV and 79.60 for NVFP4/INT8 KV. These are teacher-output similarity scores, not percentages of translation quality retained; multiple translations may be valid. INT4 evaluation reused earlier BF16/NVFP4 results, so it is not an isolated quantizer-only experiment.

The original Apache-2.0 model license is included as LICENSE. See MODEL_CHANGES.md for modifications. This is an independent quantization project, not an official Tencent release.