Hy-MT2-1.8B - A large-scale edge translation model launched by Tencent Hunyuan.

What is Hy-MT2-1.8B?

Hy-MT2-1.8B is a large-scale on-device translation model launched by Tencent Hunyuan. With only 1.8B parameters, it supports mutual translation between 33 languages, and its translation quality surpasses commercial APIs from Microsoft and Doubao on benchmarks such as FLORES-200. The model utilizes extreme quantization with 2-bit (574MB) and 1.25-bit (440MB) resolutions, enabling smooth local operation on mobile phones and PCs. It has been implemented in real-time translation of Bilibili live stream comments, achieving low-cost, high-privacy, high-quality on-device translation.

Main functions of Hy-MT2-1.8B

• 33 Language TranslationIt covers major languages such as Chinese, English, Japanese, and Korean, as well as minority languages, and supports translation between any directions.
• Local operation on the device sideThe model is compressed to 440~574MB through extremely low bit quantization, and can run in the background on ordinary mobile phones and PCs.
• Real-time translationThe response time for a single bullet comment is 500~800ms, and it has been integrated into high-frequency scenarios such as Bilibili live streaming bullet comments.
• Privacy protectionThe entire translation process is completed locally on the device, without needing to upload to the cloud or leaving the device.

Technical Principles of Hy-MT2-1.8B

• 2-bit extreme quantizationThe parameters were quantized to four discrete values {-1.5, -0.5, 0.5, 1.5} using stretch elastic quantization (SEQ), and combined with quantization-aware distillation (QAD), the translation quality was maintained at almost lossless while compressing to 574MB.
• 1.25-bit Sparse Quantization (Sherry)The core of Hunyuan's self-developed ACL 2026 Oral solution is a "fine-grained sparsity" strategy—retaining 3 ±1 and 1 0 out of every 4 parameters, with an average of only 1.25 bits required per parameter; combined with the STQ kernel designed specifically for CPU, the model size is reduced to 440MB and can run in the background.
• x86 operator adaptation optimizationIntel's team optimized vectorization, weight rearrangement, and VNNI instruction fusion for low-bit formats, resulting in a token throughput increase of approximately 2.7x and 5x for Q2_0C and STQ on mainstream Core processors, respectively, extending the ultra-low bit solution from mobile devices to PCs and edge devices.

Follow us on WeChat and reply with "open source",join inAI open source project discussion group

How to use Hy-MT2-1.8B

• Download model weightsAccess the Hugging Face repository huggingface.co/collections/AngelSlim/hy-low-bit-modelDownload the 2-bit or 1.25-bit quantized model file.
• Obtaining the reasoning frameworkClone from GitHub github.com/tencent/AngelSlim The complete code, including the extremely low bit quantization inference kernel and loading logic, is available in this repository.
• Loading and InferenceThe downloaded weights are loaded using the AngelSlim framework, and translation inference is performed on the local CPU; x86 devices can call the Intel-optimized VNNI instruction set kernel for optimal performance.

The core advantages of Hy-MT2-1.8B

• Extremely small size and extremely high massThe 1.8B parameter, after quantization with 2-bit / 1.25-bit, is only 440~574MB, yet its translation quality surpasses that of commercial APIs such as Microsoft and Doubao.
• Device-side privacy priorityThe entire translation process is completed locally on the device, without the need for internet connection for uploading, completely eliminating the risk of data leakage.
• Seamless cross-platform operationIt is compatible with both ARM mobile devices and x86 PCs, and after optimization by Intel operators, its throughput on Core processors is increased by 2.7 to 5 times.
• The real-world scenario has been implemented.It has been integrated with Bilibili live stream barrage real-time translation, with a single response time of 500~800ms, and can accurately handle popular internet slang.
• Multilingual full coverageIt natively supports translation between 33 languages, covering mainstream languages and minority languages, meeting the needs of global communication.

Hy-MT2-1.8B Project Address

• GitHub repository：https://github.com/Tencent-Hunyuan/Hy-MT2
• HuggingFace model library：https://huggingface.co/collections/AngelSlim/hy-low-bit-model

Hy-MT2-1.8B Comparison with Similar Competitors

Application scenarios of Hy-MT2-1.8B