Hy-MT2: A Family of Fast, Efficient and Powerful Multilingual Translation Models in the Wild

2025-12-29 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Advanced, extended

Summary

Hy-MT2 is a new family of multilingual translation models, available in 1.8B, 7B, and 30B-A3B (MoE) sizes, supporting 33 languages. Designed for complex real-world scenarios, these models excel in general, business, domain-specific, and instruction-following translation tasks. The 1.8B model, using AngelSlim 1.25-bit extreme quantization, requires only 440 MB storage and achieves a 1.5x inference speedup, outperforming commercial APIs like Microsoft and Doubao Translator. The larger 7B and 30B models surpass open-source competitors such as DeepSeek-V4-Pro and Kimi K2.6. Hy-MT2-30B-A3B achieves an XCOMET-XXL score of 87.47 on FLORES-200 and a GEMBA score of 95.04 on DomainMTBench, demonstrating significant improvements over its predecessor, Hy-MT1.5, in quality and efficiency across diverse deployment options including FP16, 8-bit, 4-bit, 2-bit, and 1.25-bit precisions.

Key takeaway

For ML engineers building multilingual translation systems, Hy-MT2 provides a compelling solution for diverse deployment needs. You should evaluate its 1.8B model with 1.25-bit quantization for on-device applications, as it offers a 440 MB footprint and 1.5x speedup while surpassing commercial APIs. For high-quality, instruction-following translation in professional domains, consider the 7B or 30B MoE models, which outperform leading open-source alternatives and rival closed-source systems.

Key insights

Hy-MT2 delivers high-quality, efficient multilingual translation with strong instruction-following across diverse real-world and resource-constrained scenarios.

Principles

• Scaling model size improves understanding and instruction-following.
• Mixture-of-experts balances quality and inference efficiency.
• Ultra-low-bit quantization enables efficient on-device deployment.

Method

Hy-MT2 uses MT-oriented Mid-training, Family-Centric Post-training (FCPT) with multi-teacher distillation, and AngelSlim-based quantization (up to 1.25-bit) for diverse deployment.

In practice

• Deploy 1.8B model with 1.25-bit quantization for 440 MB footprint.
• Utilize MoE architecture for high-quality, efficient large-scale translation.
• Apply rule-based pre-filtering before LLM evaluation for critical errors.

Topics

• Multilingual Translation
• Model Quantization
• Mixture-of-Experts
• Instruction Following
• On-device AI
• AngelSlim

Code references

• Tencent-Hunyuan/Hy-MT2

Best for: AI Engineer, NLP Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• Translate-R1: Cost-Aware Translation Tool Use via Reinforcement Learning — Reinforcement learning enables LLMs to intelligently decide when to use translation tools, optimizing cost and performance.
• Introducing Mellum2: A 12B Mixture-of-Experts Model by JetBrains — Mellum2 is a specialized, efficient MoE model for low-latency text and code workloads within larger AI systems.
• Transformer models for Urdu Language — Urdu's low-resource status in AI is improving through Transformer models and increased data development.
• Multilingual Coreference Resolution via Cycle-Consistent Machine Translation — Coreference resolution in low-resource languages can be significantly improved by cycle-consistent machine translation data augmentation.
• Qwen3.5-397B-A17B: the smallest Open-Opus class, very efficient model — Qwen3.5 offers multimodal, long-context capabilities with efficient sparse MoE architecture, enabling local deployment despite its large scale.
• Towards Faster Language Model Inference Using Mixture-of-Experts Flow Matching — MoE-FM enables faster, high-quality non-autoregressive language model inference by handling complex latent distributions.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.