Hy-MT1.5-1.8B-1.25bit
1.25-bit quantized Hunyuan-MT 1.8B translation model.
Last verified:
What is Hy-MT1.5-1.8B-1.25bit?
Hy-MT1.5-1.8B-1.25bit is a highly compressed (440MB) on-device translation model from Tencent's AngelSlim/Hunyuan teams, built on the 1.8B-parameter Hy-MT1.5 translation model and quantized to an effective 1.25-bit using the Sherry ternary quantization framework. It supports 33 languages, 5 dialects, and 1,056 translation directions, and is designed for offline on-device deployment on ordinary smartphones.
Hy-MT1.5-1.8B-1.25bit pricing
Pricing model: Free
Free open-source release (Hugging Face / AngelSlim)
Hy-MT1.5-1.8B-1.25bit pros
- World-class translation quality that outperforms much larger models (e.g., Tower-Plus-72B, Qwen3-32B) and commercial translation APIs despite only 1.8B parameters
- Extreme 1.25-bit Sherry quantization compresses the 3.3GB FP16 model to just 440MB with minimal accuracy loss
- Fully offline on-device deployment with custom STQ kernel enabling smooth operation on phones with limited memory; data never leaves the device
- Broad language coverage: 33 languages, 5 dialects/minority languages, and 1,056 translation directions
Hy-MT1.5-1.8B-1.25bit cons
- Only ~8x speedup over FP16 on mobile hardware, so inference speed is still constrained by device CPU
- Requires specialized STQ kernel support (recently added to llama.cpp), so ecosystem/tooling compatibility may still be limited
- Translation-focused only — not a general-purpose language model
Frequently asked questions about Hy-MT1.5-1.8B-1.25bit
What is the model size?
The 1.25-bit quantized model is 440MB, compressed from the original 3.3GB FP16 model.
How does the 1.25-bit quantization work?
It uses the Sherry framework with 3:4 fine-grained sparsity: for every 4 weights, the 3 most important are stored in 1-bit and the remaining one is zeroed, packing 4 weights into 5 bits.
How many languages does it support?
33 languages, 5 dialects/minority languages, and 1,056 translation directions.
Is there a mobile demo available?
Yes, an Android demo APK is available for download from the related Hugging Face GGUF repository.