Tencent Researchers Release Tencent HY-MT1.5: New Translation Models Including 1.8B and 7B Models Designed for Seamless Device and Cloud Use
Tencent Hunyuan researchers released HY-MT1.5, a multilingual machine translation family that targets both mobile devices and cloud systems with the same training recipe and metrics. HY-MT1.5 consists of 2 translation models, HY-MT1.5-1.8B and HY-MT1.5-7B, supports the same translation for all 33 languages with 5 ethnic and linguistic diversity, and is available on GitHub and Face Hugging under open weights.
Examples of family and referral targets
The HY-MT1.5-7B is an upgraded version of the WMT25 flagship system Hunyuan-MT-7B. It is optimized for descriptive translation and mixed language situations, and adds native support for lexicography, content translation and formatted translation.
HY-MT1.5-1.8B is compact type. It has less than a third of the parameters of the HY-MT1.5-7B but delivers the same rendering performance in the reported benchmarks. After calibration, the 1.8B model can work on peripheral devices and support real-time translation.
The calibrated HY-MT1.5-1.8B runs on devices with about 1 GB of memory and achieves an average response time of about 0.18 seconds for Chinese input of about 50 tokens, while surpassing conventional translation APIs in quality. The HY-MT1.5-7B targets server and high-end applications, where a delay of 0.45 seconds is acceptable when trading off high quality.
A complete training framework
The research team describes HY-MT1.5 as a model for a specific translation language trained through a multi-stage pipeline.
The pipe has 5 main parts:
- General pre-training: The basic model starts pre-training large-scale multilingual text with the aim of language matching. This creates shared representations across languages.
- MT oriented pre training: The model is then presented in the corresponding corporate objectives and directed towards translation. This step aligns the generation distribution with actual translation tasks instead of producing a finished text.
- Supervised fine tuning: High-quality sentence and corresponding document-level data are used to tune the model with supervised loss. This section sharpens the actual accuracy, domain coverage and specific behavior, such as ZH to EN versus EN to ZH.
- In policy sinking from 7B to 1.8B: HY-MT1.5-7B is used as a teacher in HY-MT1.5-1.8B. The research team collects nearly 1 million pieces of monolingual data in 33 languages, targets them to teachers and uses Kullback Leibler regression on student output to match the distribution of teachers. This provides a 1.8B student that inherits most of the rendering behavior of the 7B model at a much lower cost.
- Emphasize learning through rubric-based assessments: In the final stage, both models are developed with a relational policy optimization style algorithm and a rubric-based reward model. Human reviewers score translations on multiple axes such as accuracy, fluency, idiomaticity and cultural appropriateness. The award model breaks down those points and guides policy revision.
This pipeline is aimed at machine translation. It differs from the discussion-oriented LLM training by combining supervised data with a strong focus on translation, on policy classification within the translation domain and RL tuned by well-characterized translation rubrics.
Benchmark results against open and commercial systems
HY-MT1.5 is tested on Flores 200, WMT25 and the Mandarin to Young language benchmark using XCOMET-XXL and CometKiwi.
Key findings from the Table above in the report:
- In the Flores 200, the HY-MT1.5-7B achieves XCOMET-XXL scores of 0.8690 for ZH to XX, 0.9093 for EN to XX and 0.8098 for XX to XX. It outperforms specialized translation models such as the iFLYTEK translator and the Doubao translator and matches or surpasses conventional mid-sized models such as the Qwen3-235B-A22B.
- In WMT25, HY-MT1.5-7B reaches XCOMET-XXL 0.6159. This is about 0.065 higher than the Gemini 3.0 Pro and more than translation-oriented models like the Seed-X-PPO-7B and Tower-Plus-72B. The HY-MT1.5-1.8B gets a score of 0.5308, outperforming the average mid-sized models and translation systems.
- For Mandarin-to-few language pairs, HY-MT1.5-7B scores 0.6174 on XCOMET-XXL, the highest of all benchmarks including Gemini 3.0 Pro. The variance of 1.8B is up to 0.5806 and still outperforms several larger models such as DeepSeek-V3.2.
In the human evaluation on the 0 to 4 scale of Chinese to English and English to Chinese, HY-MT1.5-1.8B achieves an average score of 2.74, which is higher than the systems of Baidu, iFLYTEK, Doubao, Microsoft and Google under the same protocol.
Functional aspects of product use
The models feature three fast-moving capabilities that are essential to manufacturing systems:
- Word intervention: A quick template allows you to map a term like “混元珠 → Chaos Pearl”. Apart from the map, the model emits ambiguous transpositions. By mapping, it enforces a domain-specific term. This is important for restricted legal, medical or product content.
- Translation of information content: The second template accepts the context block and the sentence to be translated. The report shows the word “pilot” being misinterpreted as a person when there is no context. If a paragraph about a TV series is added, the model correctly interprets “pilot” as an episode.
- Translation that preserves formatting: The third template wraps the source
These are used as information formats, so they are available even if you call public weights in standard LLM stacks.
Quantization and edge transmission
HY-MT1.5-1.8B is tested with FP8 and Int4 after quantitative training using GPTQ.
Table 4 above shows:
- FP8 keeps the XCOMET-XXL score very close to the full precision model, for example 0.8379 compared to 0.8361 for ZH to XX.
- Int4 reduces the size further but presents a clear quality drop in the Flores 200.
For Hugging Face, Tencent publishes both FP8 and GPTQ Int4 variants of the HY-MT1.5-1.8B and HY-MT1.5-7B, as well as GGUF versions of local stacks. Quantization is a method that allows for a reported 1 GB memory deployment and low latency on consumer hardware.
Key Takeaways
- HY-MT1.5 is a 2-model translation family, HY-MT1.5-1.8B and HY-MT1.5-7B, which supports simultaneous translation in all 33 languages and 5 dialects or different forms, released with open weights on GitHub and Face Hugging.
- HY-MT1.5-1.8B is a distillation-based edge model that uses about 1 GB of memory and about 0.18 seconds of input latency for 50 tokens, while achieving industry-leading performance among models of similar size and outperforming most commercial translation APIs.
- The HY-MT1.5-7B is an advanced WMT25 master system that achieves about 95 percent of the Gemini 3.0 Pro in Flores 200 and surpasses it in the smaller WMT25 and Mandarin benchmarks, competing with larger open and closed models.
- Both models were trained with a complete direct pipeline of translation that includes pre-training directed at regular MT and MT, supervised fine-tuning, about policy priming and reinforced learning guided by rubric-based human assessment, which is essential for quality trade-offs and their effectiveness.
- HY-MT1.5 exposes productivity-oriented features through information, including wording, context-aware translation and format-preserving translation, and ships FP8, Int4 and GGUF variants so teams can deploy to devices or servers with standard LLM stacks.
Check it out Paper, Model weights in HF again GitHub Repo. Also, feel free to follow us Twitter and don't forget to join our 100k+ ML SubReddit and Subscribe to Our newspaper. Wait! are you on telegram? now you can join us on telegram too.
Michal Sutter is a data science expert with a Master of Science in Data Science from the University of Padova. With a strong foundation in statistical analysis, machine learning, and data engineering, Michal excels at turning complex data sets into actionable insights.
