tencent/Hunyuan-MT-7B

Hunyuan-MT-7B is Tencent’s multilingual translation model. Its basic job is to take a source passage and produce that passage in the requested target language. Tencent distinguishes this translator from Hunyuan-MT-Chimera, the family’s ensemble model that refines several candidate translations into one result.

Choose the language pair and the intended translation

Tencent describes mutual translation across 33 languages and highlights five Chinese ethnic-minority languages or dialects. The published language table includes Chinese, Traditional Chinese, English, French, Portuguese, Spanish, Japanese, Turkish, Russian, Arabic, Korean, Thai, Italian, German, Vietnamese, Malay, Indonesian, Filipino, Hindi, Polish, Czech, Dutch, Khmer, Burmese, Persian, Gujarati, Urdu, Telugu, Marathi, Hebrew, Bengali, Tamil and Ukrainian. It also explicitly lists Tibetan, Kazakh, Mongolian, Uyghur and Cantonese.

State the target language precisely, including the required writing system where relevant. A useful brief can identify whether a document needs Simplified or Traditional Chinese, and name the audience and terminology that should remain consistent. Keep the source passage intact so that the translation can be checked against the exact material supplied.

Follow the producer’s translation prompt

The manufacturer provides one prompt template for translation involving Chinese and another for language pairs that do not involve Chinese. Both name the target language, place the source text beneath the instruction and ask for the translation without an added explanation. The model has no default system prompt, so put the task instruction in the submitted message rather than relying on an unspecified default role.

For a non-Chinese pair, an original brief can begin: “Translate the following segment into French, without additional explanation.” Follow it with the source passage. When the task involves Chinese, use the corresponding Chinese template and replace its target-language placeholder. The expected artifact is the translated passage, ready for a separate review or an application’s next processing step.

Prepare segments that preserve meaning

Use coherent sentences or paragraphs instead of breaking text at arbitrary character positions. Keep the surrounding material needed to interpret pronouns, technical terms and idioms. For a product manual, include the product name, defined terms and the actual instruction together; for a support exchange, preserve the question and the answer it refers to.

If a larger document is processed in segments, keep a stable segment ID outside the text being translated and maintain the agreed terminology across the resulting passages. Ask explicitly how to handle names, measurements and placeholders, then inspect that treatment in the output. These preparation steps make the result easier to assemble and review; they do not replace checking the translated meaning.

Keep the translator and the ensemble separate

Hunyuan-MT-7B produces a translation from the source. Chimera’s documented workflow instead takes the original source plus several target-language candidates and produces a single refined translation. A workflow that uses both stages must supply those candidates to the separate ensemble model; selecting MT-7B alone does not perform that additional merge.

The producer also lists FP8 checkpoints separately. Retain the ordinary Hunyuan-MT-7B edition name when saving the request and translation, and distinguish it from an FP8 or Chimera artifact when comparing results. The later HY-MT1.5 models have their own identities and specifications.

Native inference guidance and result review

Tencent recommends temperature 0.7, top_p 0.6, top_k 20 and repetition_penalty 1.05 in its native inference example. Treat the translated content as the result to review. Check names, numbers, units, negation, dates and the relationship between sentences against the source, especially where a literal translation could lose the intended meaning.

Retain the source, target language, selected model identity and final reviewed translation together. For application text, check that placeholders and formatting still connect to the intended fields; for a document, read the joined paragraphs for consistent terminology and tone. A fluent sentence can still require correction if it changes a factual detail or an instruction.

Through Ailoot, select this exact model ID in /v1/chat/completions and supply the task in messages. Each message carries a role and its content; keep the relevant instruction and source material together. Collect the complete JSON response or all SSE segments when stream is enabled, and retain usage with the written result.