MGTV-AI/Mango-MT-9B

Mango-MT-9B: Multilingual Audiovisual Subtitle Translation

MGTV-AI's Mango-MT-9B is a 9 billion parameter model specifically engineered for multilingual audiovisual subtitle translation across 11 languages. It addresses the complexities of video subtitle translation, such as fragmented lines, timeline constraints, and maintaining plot context, which generic machine translation often struggles with.

Key Capabilities & Differentiators

• Scenario-oriented & Multilingual Optimization: Professionally optimized for 11 languages, adapting to complex industrial rules for video subtitle translation.
• Context-aware Translation: Utilizes global context modeling to ensure stable plot logic and consistent character appellation.
• Structural Robustness: Maintains original subtitle numbers and timestamps for reliable batch translation delivery.
• Superior Performance: Outperforms commercial models like DeepSeek-v4-pro, Gemini-3-pro, and GPT-5.4 across all 11 evaluated languages on the specialized Mango-SubBench, demonstrating strong multilingual translation capabilities for film and drama content.

Benchmarks

Evaluated on the Mango-SubBench, a novel benchmark for audiovisual subtitle translation, Mango-MT consistently achieves higher overall scores (S) compared to leading commercial models. For instance, it scores 0.90 for English, 0.94 for Vietnamese, and 0.92 for Spanish, surpassing competitors. While its performance on the general-purpose FLORES+ benchmark shows competitive COMET scores, its BLEU scores are generally lower than some commercial models, highlighting its specialized optimization for audiovisual content.

Good For

• Industrial-scale video subtitle translation: Ideal for mass production of subtitles where timeline compliance, consistent semantics, and plot context are critical.
• Multilingual content localization: Translating film, drama, and other long-form video content into Malay, Thai, English, Indonesian, Vietnamese, Russian, French, Japanese, Korean, Spanish, and Arabic.