Hy-MT2-1.8B - Tencent Hunyuan's On-Device Translation Large Model

What is Hy-MT2-1.8B

Hy-MT2-1.8B is an on-device translation large model released by Tencent Hunyuan. With only 1.8B parameters, it supports translation among 33 languages, and its translation quality surpasses commercial APIs such as Microsoft and Doubao on benchmarks like FLORES-200. Through extreme quantization at 2-bit (574MB) and 1.25-bit (440MB), the model can run smoothly locally on phones and PCs. It has already been deployed for real-time translation of Bilibili live-stream danmaku, enabling low-cost, high-privacy, high-quality on-device translation.

Main Features of Hy-MT2-1.8B

• 33-language mutual translation: Covers mainstream languages such as Chinese, English, Japanese, and Korean as well as minority languages, and supports translation in any direction.
• On-device local operation: Through extremely low-bit quantization, the model is compressed to 440~574MB, and can reside in the background on ordinary phones and PCs.
• Real-time translation: A single danmaku responds in 500~800ms, and it has been integrated into high-frequency scenarios such as Bilibili live-stream danmaku.
• Privacy protection: The entire translation process is completed locally on the device, with no need to upload to the cloud, and data never leaves the device.

Technical Principles of Hy-MT2-1.8B

• 2-bit extreme quantization: It uses Stretched Elastic Quantization (SEQ) to quantize parameters into four discrete values {-1.5, -0.5, 0.5, 1.5}. Combined with Quantization-Aware Distillation (QAD), it maintains almost lossless translation quality while compressing to 574MB.
• 1.25-bit sparse quantization (Sherry): Hunyuan's self-developed ACL 2026 Oral solution. Its core is a "fine-grained sparsity" strategy - among every 4 parameters, 3 are kept as ±1 and 1 is set to 0, requiring an average of only 1.25-bit per parameter. Together with the STQ kernel designed specifically for CPU, the model size is reduced to 440MB and can run resident in the background.
• x86 operator adaptation and optimization: The Intel team optimized vectorization, weight reordering, and VNNI instruction fusion for low-bit formats, increasing the token throughput of Q2_0C and STQ on mainstream Core processors by about 2.7x and 5x respectively, extending extremely low-bit solutions from mobile devices to PCs and edge devices.

Follow WeChat and reply "open source" to joinAI Open Source Project Exchange Group

How to Use Hy-MT2-1.8B

• Download model weights: Visit the Hugging Face repository huggingface.co/collections/AngelSlim/hy-low-bit-modeland download the model files after 2-bit or 1.25-bit quantization.
• Obtain the inference framework: Clone from GitHub github.com/tencent/AngelSlim The complete code. This repository contains the inference kernel and loading logic for extremely low-bit quantization.
• Loading and inference: Use the AngelSlim framework to load the downloaded weights and perform translation inference on a local CPU; x86 devices can call Intel-optimized VNNI instruction set kernels for optimal performance.

Core Advantages of Hy-MT2-1.8B

• Extremely small size, extremely high quality: After 2-bit / 1.25-bit quantization, the 1.8B parameters are only 440~574MB, yet the translation quality surpasses commercial APIs such as Microsoft and Doubao.
• On-device privacy first: The entire translation process is completed locally on the device, with no need for network upload, completely eliminating the risk of data leakage.
• Seamless cross-platform operation: It is compatible with both ARM mobile devices and x86 PCs. After Intel operator optimization, throughput on Core processors increases by 2.7~5x.
• Already deployed in real scenarios: Real-time translation of Bilibili live stream comments is now integrated, with 500–800 ms response per message, accurately handling internet slang.
• Full multilingual coverage: Native support for translation among 33 languages, covering major languages as well as minority languages, meeting global communication needs.

Project address of Hy-MT2-1.8B

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

Comparison of Hy-MT2-1.8B with similar competing products

Application scenarios of Hy-MT2-1.8B