Tencent Hyuan Open-Sports Hiyuan-MT-7B and Huyuan-Mt-Chima-chima: Models to translate multiple languages

Introduction

Tencent's Header's Team Team out Hyyuan-MT-7B (translation model) and Hyyuan-Mt-chimera-7b (enbble model). Both models are specifically designed for multilingual machine translation and is presented in conjunction with tencent participation in the WMT2025 Normal Translation of Shared MachineWhen Hyuuan-MT-7b is first placed on 30 from 31 two languages.

Overview of the model

Hyyuan-MT-7B

• A The 7B parameter Charess Model model.
• Is supporting Translation Mutual in 33 languagesincluding Languages ​​of Chinese Raciously Like Tibetan, Musongolian, Uyghur, and Kazakh.
• It is for both Translation tasks with higher and lower resourcesGetting the State-The-Art State Results between the size of the unique size.

Hyyuan-Mt-chimera-7b

• A The combined model with strong weaknesses.
• It includes many translations during translation and produces an analyzed interpretation using learning and integrated strategies.
• Represents the The original source model of this typeImproving the quality of translation without a single system.

Outline of training

The models are trained using a Five Stage Framework Designed for translation activities:

1. The average of previous
   • 1.3 Millions of tokens cover 112 languages ​​and languages.
   • Corporal Corpay is tested for a number of information, authenticity, and writing style.
   • The diversity is stored by disciplinary, industry, and additional marking plans.
2. Previous Training of MT-SUSE
   • MONOKAL CORPORA from MC4 and Oscar, filtered using FastText (Language ID), Minlsh (Deleting), and Decupleyity Filting).
   • Parallel Corpora from Opus and Paracrawl, filtered with cometkiwi.
   • The multiple training data repetition (20%) to avoid disaster loss.
3. To direct the beauty of directive (sft)
   • Stage I: ~ 3m in combined pairs (flores-200, WMT test sets, selected Mandarin-Minity data, in pairs of activities, teaching data).
   • SECTION II: ~ 268K in the highest quality pairs of the Cometkiwi, Camp) and handle.
4. Emphasizing reading (rl)
   • Algorithm: Salmon.
   • Rewards:
     • XCOMET-XXL and DEEPSEEK-V3-0324 Making Quality.
     • Terminology – Knowledge Rewards (Tat-R1).
     • Repeated penalties to avoid flowing out.
5. Weak-to-Trong RL
   • Many effects of cash flow and compiled with the removal from the reward
   • Used in Hyyuan-Mt-chimera-7bImproving translation flexibility and reducing repetitive mistakes.

Benchmark results

Automatic testing

• WMT24Pp (English⇔xx): Huyuan-MT-7B reached 0.8585 (XCOMET-XXL)Exceeding large models such as Gemini-2.5-Pro ​​(0.8250) and Claudude-Sonnet-4 (0.8120).
• Florer-200 (33 Languages, 1056 pairs): Huyuan-MT-7B found points 0.8758 (XCotet-XXL)Open source of the open source include QWEN3-32B (0.7933).
• Mandarin⇔mine Languages: Points 0.6082 (xcotet-xxl)higher than Gemini-2.5-Pro ​​(0.5811), indicating important improvements in lower settings.

Confidential effects

• Outperforms Google translator 15-65% in all test stages.
• Special Special Translation models are like Tower-plus-9b including Set-X-PPO-7B Despite having few parameters.
• Chimera-7b Adds ~ 2.3% Upgrade to Flores-200, especially in Chinese and English-Chinese translation.

Personal examination

A custom set (covering social, medical, medical, law, law) Compare Huyuan-MT-7B in the country models:

• Hyyuan-MT-7B: Avg. 3.189
• Gemini-2.5-pro: Avg. 3.23
• Deepseek-v3: Avg. 3.219
• Google Translate: Avg. 2.344

This shows that Hyyuan-MT-7B, although under 7b parameters, it approaches the quality of the largest model models.

Subject lessons

The report highlights several cases of the real world:

• Customs References: Ethical “小红薯” as platform “rednote” unlike Google Translate “sweet potatoes.”
• Idioms: Translating “kills me” as “你真要把我笑死了” (express entertainment), to avoid unpleasant translation.
• Medical Conditions: Translated “stones of Urig acid kids” accurately, while baseline produce unwanted results.
• Small languagesIn Kazakh and Tibetan, Hyyuan-MT-7b produces united versions, where the foundations failed or removes unrealistic text.
• Chimera enhancements: It adds an improvement in the Gargon Gargon, stability, and sport termology.

Store

Tencent Issue Hyyuan-MT-7B including Hyyuan-Mt-chimera-7b He establishes a new standard of the open source. By combining a carefully designed training framework with exclusive focus on The translation of the lower and lessModels reach quality in a robber or exceeding large source systems. The launch of these 2 models provide the AI ​​and the tools that apply to many translation and distribution languages.

Look Page, GitHub Page, and model in face view. All credit for this study goes to research for this project. Feel free to look our GITHUB page for tutorials, codes and letters of writing. Also, feel free to follow it Sane and don't forget to join ours 100K + ml subreddit Then sign up for Our newspaper.

Asphazzaq is a Markteach Media Inc. According to a View Business and Developer, Asifi is committed to integrating a good social intelligence. His latest attempt is launched by the launch of the chemistrylife plan for an intelligence, MarktechPost, a devastating intimate practice of a machine learning and deep learning issues that are clearly and easily understood. The platform is adhering to more than two million moon visits, indicating its popularity between the audience.