Tencent’s Hunyuan team has released Hunyuan-MT-7B (a translation model) and Hunyuan-MT-Chimera-7B (an ensemble model). The two systems are focused on multilingual machine translation and debuted as Tencent took part in the WMT2025 General Machine Translation shared task, where Hunyuan-MT-7B finished first in 30 of 31 language pairs.
Hunyuan-MT-7B is a 7B-parameter translation model that supports bidirectional translation across 33 languages, including Chinese ethnic minority languages such as Tibetan, Mongolian, Uyghur, and Kazakh. The model was tuned for high-resource and low-resource scenarios and produced state-of-the-art results for models of comparable size.
Hunyuan-MT-Chimera-7B implements an integrated weak-to-strong fusion strategy. At inference it merges multiple translation candidates, then refines the aggregated output with reinforcement learning and reward-driven aggregation. This design is presented as the first open-source translation model using that fusion approach, yielding quality above single-system outputs.
Training followed a five-stage framework tailored to translation. The pre-training mix totaled 1.3 trillion tokens spanning 112 languages and dialects. Multilingual corpora were evaluated for knowledge value, authenticity, and writing style. Diversity was preserved through disciplinary, industry, and thematic tagging systems.
Monolingual sources included mC4 and OSCAR, filtered with fastText for language ID, minLSH for deduplication, and KenLM for perplexity-based filtering. Parallel corpora came from OPUS and ParaCrawl and were filtered with CometKiwi. The pipeline replays 20% of general pre-training data to curb catastrophic forgetting.
Stage I used roughly 3 million parallel pairs drawn from Flores-200, WMT test sets, curated Mandarin–minority data, synthetic pairs, and instruction-tuning examples. Stage II selected about 268,000 high-quality pairs through automated scoring with CometKiwi and GEMBA, followed by manual verification.
The optimization algorithm applied is GRPO. Reward signals include XCOMET-XXL and DeepSeek-V3-0324 to score translation quality, a terminology-aware reward labeled TAT-R1, and repetition penalties to reduce degenerate outputs. Multiple candidate translations are generated and combined via reward-based aggregation; that mechanism is applied in Hunyuan-MT-Chimera-7B to improve robustness and cut down on repetitive errors.
Benchmark results show strong performance. On WMT24pp (English⇔XX) Hunyuan-MT-7B scored 0.8585 (XCOMET-XXL), ahead of larger systems such as Gemini-2.5-Pro (0.8250) and Claude-Sonnet-4 (0.8120). On FLORES-200 (33 languages, 1,056 pairs) Hunyuan-MT-7B reached 0.8758 (XCOMET-XXL), outperforming open-source baselines including Qwen3-32B (0.7933). For Mandarin⇔minority languages the model scored 0.6082 (XCOMET-XXL), above Gemini-2.5-Pro’s 0.5811 and representing notable gains in low-resource settings.
Across evaluation categories Hunyuan-MT-7B outperforms Google Translator by 15–65%. The 7B model outperforms specialist translation systems such as Tower-Plus-9B and Seed-X-PPO-7B and uses fewer parameters. Hunyuan-MT-Chimera-7B provides roughly a 2.3% lift on FLORES-200, with the largest gains in Chinese⇔other and non-English⇔non-Chinese directions.
A custom evaluation set covering social, medical, legal, and internet domains compared Hunyuan-MT-7B with state-of-the-art offerings. Average scores were: Hunyuan-MT-7B 3.189; Gemini-2.5-Pro 3.223; DeepSeek-V3 3.219; Google Translate 2.344. The results indicate Hunyuan-MT-7B, smaller at 7B parameters, approaches the quality of much larger proprietary models.
The report includes practical examples that illustrate model behavior. Cultural references: "小红薯" is translated as "REDnote," while Google Translate returns "sweet potatoes." Idioms: "You are killing me" maps to "你真要把我笑死了" to convey amusement rather than a literal reading. Medical terms: "uric acid kidney stones" are rendered precisely; some baseline systems produce malformed outputs. Minority languages: for Kazakh and Tibetan Hunyuan-MT-7B generates coherent translations where other systems fail or output nonsense. Chimera gains add improvements for gaming jargon, intensifiers, and sports terminology.
Tencent’s publication of Hunyuan-MT-7B and Hunyuan-MT-Chimera-7B establishes a new reference point for open-source translation research. The release pairs a carefully structured training pipeline with a focused effort on low-resource and minority-language translation, producing results comparable to or exceeding larger closed-source alternatives. The two models provide the AI research community with accessible, high-performance tools for multilingual translation research and deployment.