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

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

Hy-MT1.5-1.8B-1.25bit is an ultra-compressed offline translation model for mobile devices, developed by Tencent's Hunyuan team. Based on the Hunyuan translation model Hy-MT1.5-1.8B (1.8 billion parameters), it utilizes innovative Sherry sparse ternary quantization technology to compress the model to a mere 440MB. It supports 33 languages, 5 dialects/minority languages, and 1056 translation directions. It can run locally on a mobile phone without an internet connection, and its translation quality surpasses that of mainstream commercial systems such as Google Translate.

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

• Offline translation between 33 languagesIt covers major languages such as Chinese, English, French, Japanese, Arabic, and Russian, as well as minority languages such as Tibetan and Mongolian, with a total of 1,056 translation directions.
• Run locally on mobile deviceAfter downloading the model, it can be translated completely offline, without requiring a network connection and consuming no data.
• Background word retrieval and translationThe accompanying demo supports a "background word-capture mode," allowing you to call up translations at any time while browsing web pages or checking emails.
• Real-time text translationQuickly generate translation results after you enter or paste text, and support copying and output.
• Multilingual direction switchingSupports bidirectional translation switching between any language pairs.

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

• Basic ModelBased on Tencent's Hy-MT1.5-1.8B (1.8B parameters) large-scale translation model, this model is built through a multi-stage training framework of "general pre-training → MT-oriented pre-training → supervised fine-tuning → online distillation → reinforcement learning".
• Sherry sparse ternary quantization (1.25-bit)A fine-grained sparse strategy is adopted, where 3 out of every 4 parameters are stored using 1 bit (with a value of -1 or 1), and 1 is set to 0, requiring only 1.25 bits per parameter on average.
• STQ dedicated kernelTencent's sparse ternary quantization inference kernel, designed specifically for mobile CPUs, is perfectly compatible with the SIMD instruction set, enabling efficient decoding.
• Quantization-Aware Training (QAT)By optimizing the distribution alignment and offset during training, the accuracy loss caused by ultra-low bit quantization can be mitigated.
• GRPO reinforcement learningThe Group Relative Policy Optimization algorithm is used, combined with multi-dimensional scoring criteria (accuracy, fluency, consistency, cultural compatibility, and readability) to optimize translation quality.

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

• Download and installDownload and install 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.
• Startup loadingThe first time you open the app, it will automatically load a 440MB model. Once loaded, it will be ready to use.
• Text translationEnter or paste text, click Translate for offline translation; click the language bar at the bottom to switch the translation direction.
• Background word retrievalIn other apps, long-press to select text, choose "Offline Translation," and the result will be displayed directly in a floating window without needing to switch apps.

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

• PublisherTencent Hunyuan Team
• Model parameters1.8B
• Quantization accuracy1.25-bit (Sherry sparse ternary quantization)
• Model volume440MB
• Supported PlatformsAndroid (Demo version), iOS official version will be supported later.
• Open source licenseThe model weights, code, and technical reports are all fully open source.
• systemAndroid phones (Demo version does not currently support iOS)
• MemoryRecommended RAM: 8GB or more (demonstration devices are Snapdragon 865/888/7+Gen2, 8GB/16GB RAM)
• storageApproximately 500MB of space is reserved for downloading the model.
• networkThe first time you download the model, you need an internet connection; afterwards, you can use it completely offline.

The core advantages of Hy-MT1.5-1.8B-1.25bit

• Extremely lightweightThe model is only 440MB, while the original FP16 version is 3.3GB, with a compression ratio of about 7.5 times. It can run in the background on ordinary mobile phones.
• High translation qualityOn benchmarks such as Flores-200 and WMT25, the 1.8B original model has surpassed commercial APIs such as Google Translate, Baidu Translate, and Microsoft Translate, as well as larger open-source models such as Tower-Plus-72B and Qwen3-32B; the 1.25-bit quantized version maintains a very small size while keeping quality loss manageable.
• Fast reasoning speedWith the STQ kernel, the translation speed on Snapdragon 888 phones is about 8 times faster than the FP16 version (1.9 seconds vs 58.2 seconds).
• Completely offline with zero privacyCompletely local processing, no data upload required, no subscription fees, one-time download for permanent use.
• Hardware compatibilityThe 1.25-bit scheme is compatible with the SIMD instruction set and can run smoothly on all Android models.

Hy-MT1.5-1.8B-1.25bit project address

• HuggingFace model libraryhttps://huggingface.co/AngelSlim/Hy-MT1.5-1.8B-1.25bit
• arXiv technical paper: https://arxiv.org/pdf/2512.24092

Hy-MT1.5-1.8B-1.25bit: Comparison with similar competing products

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

• Traveling abroadReal-time translation of menus, road signs, and conversations in environments without network access (remote areas, airplanes, subways), solving the pain point of unstable signals overseas.
• Business OfficeOffline processing of international emails and contracts ensures that sensitive business information is processed entirely locally, eliminating the risk of data upload.
• Communication with ethnic minoritiesIt supports translation between Chinese and other languages, including Tibetan, Mongolian, and Uyghur, to facilitate daily communication and government affairs in border areas.
• academic researchLocal access to foreign language literature and papers, with instant word translation, protecting academic privacy and intellectual property rights.
• Daily ReadingOffline browsing of foreign language websites, social media, and e-books allows you to access high-quality translations without using data.