Hy-MT2-7B Guide: Glossaries, Delimiters and Structured Translation

The hard part of machine translation in production is rarely fluency. It is everything around it: a product name that must never be translated, a %s that must survive, a JSON file where only the values should change, a brand voice that is formal in German and casual in Spanish. Hy-MT2-7B is interesting because Tencent built prompt templates for exactly those cases and trained the model to follow them.

It is the middle model of the Hy-MT2 family Tencent open-sourced on 21 May 2026, between Hy-MT2-1.8B and Hy-MT2-30B-A3B. All three are Apache 2.0. The 7B is the one most teams can realistically self-host on a single ordinary GPU.

From Hunyuan-MT-7B to Hy-MT2-7B

The 7B has two predecessors worth knowing about. Tencent open-sourced Hunyuan-MT-7B on 1 September 2025, along with Hunyuan-MT-Chimera-7B, an ensemble model that merges several candidate translations into one. Tencent reported that Hunyuan-MT-7B placed first in 30 of the 31 language categories it entered at WMT25. HY-MT1.5, in 1.8B and 7B sizes, followed on 30 December 2025.

Hy-MT2 keeps the 33-language scope and the 1.8B and 7B size points, and adds two things. First, a larger tier: a 30B-A3B MoE aimed at quality. Second, a much heavier emphasis on instruction following. The release introduced IFMTBench, a benchmark for how well translation models obey instructions, and the technical report frames the family around complex real-world scenarios rather than clean sentence pairs.

Tencent reports that the 7B and 30B-A3B outperform DeepSeek V4 Pro and Kimi K2.6 in fast-thinking mode on its evaluation. That is a vendor claim with results published as a chart, so treat it as a reason to test rather than a settled fact.

Seven templates, and why to use them verbatim

The model card publishes prompt templates in Chinese and English for seven workflows:

• Default — translate and return only the translation, with no commentary.
• Terminology — supply reference term pairs before the text, so the model uses your glossary.
• Style — require the output to match a stated target style.
• Personalisation — list user preferences as numbered requirements.
• Delimiters — require the output to keep the same number of delimiters in the same positions.
• Structured data — translate only user-facing text inside JSON or code strings, leaving keys and placeholders unchanged.
• Contextual — add background information to steer the translation.

Copy these from the model card exactly rather than paraphrasing them. A specialist model trained on fixed instruction formats tends to follow them most reliably in the wording it was trained on. There is no default system prompt, so the template carries all of the instruction.

The delimiter template deserves particular attention. A common pattern for translating many short strings is to join them with a separator, translate once and split the result. That breaks the moment a model merges or drops a separator. A template that explicitly requires delimiter count and position to be preserved, on a model trained for it, makes that batching trick far safer. Still verify the count on every response and retry on a mismatch.

Running it yourself

The 7B is a dense model, listed on Hugging Face at roughly 8B parameters in BF16. Tencent's serving example uses one GPU:

vllm serve tencent/Hy-MT2-7B --tensor-parallel-size 1

SGLang works with --tp 1, and Transformers needs version 5.6.0 or later. Official FP8 and GGUF builds are published, so it also runs under llama.cpp on smaller hardware, though the card notes the GGUF depends on a quantisation kernel Tencent contributed to llama.cpp, so use a recent build. Best model for self-hosting covers how to size a GPU for a model of this class.

Use the recommended sampling settings: temperature 0.7, top_p 0.6, top_k 20, repetition penalty 1.05, with generation capped at 4,096 tokens. vLLM accepts top_k and repetition_penalty as extra parameters on its OpenAI-compatible endpoint, so a request body looks like this:

{
  "model": "tencent/Hy-MT2-7B",
  "messages": [{"role": "user", "content": "...template with your text..."}],
  "temperature": 0.7,
  "top_p": 0.6,
  "top_k": 20,
  "repetition_penalty": 1.05,
  "max_tokens": 4096
}

Hosted listings, including OpenRouter, give every Hy-MT2 size an 8,192-token context, so plan on chunking long documents and resending the glossary with each chunk.

The language list, read carefully

The report and model card introduction say 33 languages. The Hugging Face metadata says 36, and the supported-language table has 38 rows. The difference is variants: the table lists Traditional Chinese and Cantonese separately, for example. Beyond the major European and Asian languages it includes Khmer, Burmese, Tibetan, Kazakh, Mongolian and Uyghur.

That breadth is a reason to test the long tail properly. A model can be strong on English to French and much weaker on Burmese to Kazakh, and an average hides that. Evaluate the specific pairs you serve, in both directions, since quality is rarely symmetric.

Picking the 7B over its siblings

On a hosted API, OpenRouter lists the 7B at $0.074 input and $0.295 output per million tokens at the time of writing, the same as the 30B-A3B. If you are only ever calling an API, the 30B-A3B is the more sensible default at that price.

The 7B earns its place when you self-host. It is dense, fits on one GPU, has official GGUF builds and is supported for full-parameter and LoRA fine-tuning through LLaMA-Factory. If you want to adapt a translator to your own domain and run it on your own hardware, this is the size to start with. Drop to the 1.8B only when the model has to run on the end user's device.

For the general craft of translation prompts, including register and placeholders, see LLM translation: placeholders, glossaries and register.