Hy-MT1.5-1.8B-1.25bit - Tencent Hunyuan's open-source on-device offline translation model

What is Hy-MT1.5-1.8B-1.25bit

Hy-MT1.5-1.8B-1.25bit is an extremely quantized and compressed on-device offline translation large model released by the Tencent Hunyuan team. The model is built on the Hunyuan translation large model Hy-MT1.5-1.8B(1.8 billion parameters), and through the innovative Sherry sparse ternary quantization technology, the model is compressed to only 440MB. It supports 33 languages, 5 dialects/minority languages, and 1056 translation directions, can run locally on mobile phones without an internet connection, and its translation quality surpasses mainstream commercial systems such as Google Translate.

Main features of Hy-MT1.5-1.8B-1.25bit

• Offline mutual translation among 33 languages: Covers mainstream languages such as Chinese-English, French-Japanese, Arabic, and Russian, as well as minority languages such as Tibetan and Mongolian, with a total of 1056 translation directions.
• Local operation on mobile phones: After downloading the model, translation can be performed completely offline, without an internet connection, and without consuming data.
• Background word capture translation: The accompanying Demo supports a "background word capture mode", allowing translation to be invoked at any time while browsing web pages or checking emails.
• Real-time text translation: Quickly generates translation results after inputting or pasting text, and supports copying the output.
• Multilingual direction switching: Supports bidirectional translation switching between any language pair.

Technical principles of Hy-MT1.5-1.8B-1.25bit

• Base model: Based on the Tencent Hunyuan translation large model Hy-MT1.5-1.8B (1.8B parameters), this model is built through a multi-stage training framework of "general pretraining -> MT-oriented pretraining -> supervised fine-tuning -> online distillation -> reinforcement learning".
• Sherry sparse ternary quantization (1.25-bit): Adopts a fine-grained sparsity strategy, in which 3 out of every 4 parameters are stored with 1-bit (values -1 or 1), and 1 is set to 0, requiring an average of only 1.25 bits per parameter.
• STQ dedicated kernel: A sparse ternary quantization inference kernel designed by Tencent specifically for mobile phone CPUs, perfectly adapted to SIMD instruction sets for efficient decoding.
• Quantization-aware training (QAT): Through distribution alignment and offset optimization during training, it alleviates the accuracy loss caused by ultra-low-bit quantization.
• GRPO reinforcement learning: Adopts the Group Relative Policy Optimization algorithm, combined with multi-dimensional scoring criteria (accuracy, fluency, consistency, cultural adaptability, readability) to optimize translation quality.

How to use Hy-MT1.5-1.8B-1.25bit

• Download and install: Download the Demo APK from the ModelScope community https://modelscope.cn/models/AngelSlim/Hy-MT1.5-1.8B-1.25bit-GGUF/resolve/master/Hy-MT-demo.apk or HuggingFace and install it.
• Launch and load: On first opening, the 440MB model is automatically loaded, and it can be used after loading is complete.
• Text translation: Input or paste text and click Translate for offline translation; click the language bar at the bottom to switch the translation direction.
• Background word capture: In other apps, long-press to select text and choose "offline translation"; the floating window directly displays the result without jumping to another page.

Key information and usage requirements of Hy-MT1.5-1.8B-1.25bit

• Publisher: Tencent Hunyuan Team
• Model parameters：1.8B
• Quantization precision: 1.25-bit (Sherry sparse ternary quantization)
• Model size：440MB
• Supported platforms: Android (Demo version), iOS official version will be supported later
• Open Source License: Model weights, code, and technical report are fully open-sourced
• System: Android phones (the Demo version does not support iOS yet)
• Memory: 8GB or above is recommended (demo devices were Snapdragon 865/888/7+Gen2 with 8GB/16GB memory)
• Storage: Reserve about 500MB of space to download the model
• Network: An internet connection is required to download the model for the first time; after that, it can be used completely offline

Core advantages of Hy-MT1.5-1.8B-1.25bit

• Extreme lightweight design: The model is only 440MB, while the original FP16 version is 3.3GB, a compression ratio of about 7.5x, so it can stay resident in the background on ordinary phones.
• High translation quality: On benchmarks such as Flores-200 and WMT25, the original 1.8B model has surpassed commercial APIs such as Google Translate, Baidu Translate, and Microsoft Translator, as well as larger open-source models such as Tower-Plus-72B and Qwen3-32B; the 1.25bit quantized version keeps a very small size while keeping quality loss controllable.
• Fast inference speed: With the STQ kernel, translation speed on a Snapdragon 888 phone is about 8x faster than the FP16 version (1.9 seconds vs 58.2 seconds).
• Fully offline with zero privacy risk: Completely local processing, no need to upload any data, no subscription fees, one download for permanent use.
• Hardware-friendly adaptation: The 1.25bit solution is friendly to SIMD instruction sets, and Android phones across the board can run it smoothly.

Project address of Hy-MT1.5-1.8B-1.25bit

• HuggingFace Model Repository：https://huggingface.co/AngelSlim/Hy-MT1.5-1.8B-1.25bit
• arXiv Technical Paper：https://arxiv.org/pdf/2512.24092

Comparison of Hy-MT1.5-1.8B-1.25bit with similar competitors

Application scenarios of Hy-MT1.5-1.8B-1.25bit

• Overseas travel: Real-time translation of menus, road signs, and conversations in no-network environments (remote areas, airplanes, subways), solving the pain point of unstable overseas signal.
• Business office: Offline processing of foreign-related emails and contract documents; sensitive business information is processed entirely locally, eliminating the risk of data upload.
• Ethnic minority communication: Supports translation between Tibetan, Mongolian, Uyghur, and other minority languages and Chinese, serving daily communication and government affairs handling in border areas.
• Academic research: Read foreign-language literature and papers locally, look up word translations at any time, and protect academic privacy and intellectual property.
• Daily reading: Browse foreign-language web pages, social media, and e-books offline, and obtain high-quality translations without needing mobile data.