Hunyuan-MT-7B Multilingual Translation Model

• Hunyuan-MT-7B is an open-source multilingual translation model featuring 7 billion parameters for accurate bidirectional translation across 33 languages.
• It employs a multi-stage training methodology with general and MT-oriented pre-training, supervised fine-tuning, and reinforcement learning to boost performance on both high- and low-resource language pairs.
• The associated Chimera variant introduces a 'slow thinking mode' that uses multi-candidate decoding and RL-driven fusion to refine test-time translations for enhanced quality.

Hunyuan-MT-7B is an open-source, 7-billion parameter multilingual translation model specifically architected for robust bidirectional translation across 33 major languages, with a pronounced emphasis on Mandarin and its translation to and from several ethnic minority languages and dialects. Its design serves as a foundation for the Hunyuan-MT-Chimera-7B variant, which implements a novel "slow thinking mode" for test-time translation refinement. Hunyuan-MT-7B leverages a multi-stage training regime incorporating general and MT-oriented pre-training, supervised fine-tuning, and reinforcement learning, establishing competitive performance benchmarks in both high- and low-resource language translation tasks.

1. Model Architecture and Design

Hunyuan-MT-7B utilizes a modern Transformer-style architecture optimized for large-scale multilingual machine translation. With 7 billion parameters, it is engineered to handle both mainstream (Mandarin, English, Japanese) and low-resource/minority languages (Kazakh, Uyghur, Mongolian, Tibetan). The architecture integrates data processing and fine-tuning strategies tailored for linguistic peculiarities that arise in low-resource translation pairs.

The model serves as the backbone for Hunyuan-MT-Chimera-7B, which adopts a multi-candidate decoding strategy ("slow thinking mode"). In Chimera, Hunyuan-MT-7B generates multiple translations under varied decoding hyperparameters, and a learned fusion mechanism synthesizes these into a single output. This process, described as "weak-to-strong" synthesis, surpasses conventional chain-of-thought (CoT) refinement approaches by exploiting contextual diversity across candidate outputs.

2. Training Methodology

Hunyuan-MT-7B is trained via a holistic multi-phase methodology:

• General Pre-training: The initial phase leverages a multilingual corpus (~1.3 trillion tokens) spanning 112 languages/dialects. A proprietary quality assessment module evaluates dimensions such as Knowledge Value, Authenticity, and Writing Style. Data balance is achieved through a tripartite taxonomic framework involving disciplinary, industry, and thematic tags, providing broad linguistic coverage.
• MT-Oriented Pre-training: Data is sourced from monolingual corpora (mC4, OSCAR) and bilingual resources (OPUS, ParaCrawl). Preprocessing steps include language identification, deduplication, and quality scoring (CometKiwi). A RegMix-inspired method establishes an optimal data mixture ratio, and 20% of the original corpus is replayed to mitigate catastrophic forgetting.
• Supervised Fine-Tuning (SFT): SFT occurs in two stages: the first using ~3 million parallel translation pairs, including human-annotated sets for minority languages and instruction data; the second uses a more selective, high-fidelity dataset (~268,000 pairs), with additional human review for sample consistency.
• Reinforcement Learning (RL) and Weak-to-Strong RL: Post-SFT, RL is applied using a composite reward function: overall quality (XCOMET-XXL, DeepSeek-V3-0324), terminology-aware word alignment (TAT-R1), and repetition penalties. The weak-to-strong RL phase involves generating multiple candidates and utilizing RL-driven fusion for a single superior output—the key mechanism behind "slow thinking" test-time scaling.

3. Performance Evaluation

Hunyuan-MT-7B is evaluated on both automatic and human-centric metrics:

• Automatic: The model leverages XCOMET-XXL and CometKiwi, which demonstrate high concordance with human assessments.
• Human: Evaluations use a 0–4 scale based on accuracy, fluency, and idiomaticity.

Benchmark results show Hunyuan-MT-7B achieving superior scores to similarly sized translation models, outperforming Gemini-2.5-Pro and Claude-Sonnet-4 on the WMT24pp development set (XCOMET-XXL of 0.8585). For Mandarin↔Minority language translation, the model yields substantial improvements relative to existing baselines.

4. The Chimera Variant and “Slow Thinking Mode”

Hunyuan-MT-Chimera-7B is a test-time inference variant that operationalizes the "slow thinking mode." This two-stage process involves:

1. Hunyuan-MT-7B generating a portfolio of diverse candidate translations via manipulated decoding conditions.
2. A fusion mechanism (RL-based aggregation) combines these candidates into one refined output.

This approach extends inference time and facilitates deeper contextual and terminological consideration, leading to more coherent, robust translations. Conventional single-pass decoders cannot replicate this contextual synthesis, especially in linguistically complex or low-resource settings.

5. Language Support Spectrum

Hunyuan-MT-7B supports bidirectional translation for 33 languages. While optimized for Chinese, English, Japanese, and French, it demonstrates particular strength in low-resource scenarios—Mandarin↔Minority language pairs (Kazakh, Uyghur, Mongolian, Tibetan). Model fine-tuning leverages human-annotated and synthetic data to enhance minority and dialect translation, addressing underrepresentation and inclusivity in MT systems.

6. Empirical Results

Comprehensive evaluations on Flores-200 and WMT24pp confirm the model’s statistical dominance:

$\begin{array}{lccccccc} \textbf{Model} & \textbf{Metrics} & \text{ZH}\Rightarrow\text{XX} & \text{XX}\Rightarrow\text{ZH} & \text{EN}\Rightarrow\text{XX} & \text{XX}\Rightarrow\text{EN} & \text{XX}\Rightarrow\text{XX} & \text{WMT24pp} \ \hline \text{Hunyuan‑MT‑7B} & \text{XCOMET-XXL} & 0.8758 & 0.8528 & 0.9112 & 0.9018 & 0.7829 & 0.8585 \ \text{Hunyuan‑MT‑Chimera‑7B} & \text{XCOMET-XXL} & 0.8974 & 0.8719 & 0.9306 & 0.9132 & 0.8268 & 0.8787 \ \end{array}$

In the WMT2025 shared task, Hunyuan-MT models ranked first for 30 out of 31 language pairs. Mandarin↔Minority language benchmarks demonstrate a marked increase (~4.7%) over the next best system, evidencing robustness across both high- and low-resource languages.

7. Significance and Implications

Hunyuan-MT-7B advances multilingual machine translation with rigorous architectural and training innovations. Its tailored approach to low-resource and minority languages—as enabled by specialized data pipelines and "slow thinking" inference strategies—positions it as a state-of-the-art solution for translation tasks spanning a broad linguistic spectrum. The open-source release and superior benchmark results facilitate further research and application in multilingual translation, underpinning inclusivity and linguistic diversity in contemporary MT system design.