Tencent’s Hy‑MT2 Translation Models Explained

Language Coverage

Hy‑MT2 supports bidirectional translation across 33 languages.

In addition, Tencent reports that the system also supports translation involving five Chinese ethnic languages or dialect variants, expanding its usability in multilingual regions within China.

Although public announcements confirm the 33‑language coverage, the complete official language list is not fully enumerated in the available summaries of the paper. What is clear is that the system focuses strongly on Chinese‑centric multilingual translation scenarios alongside widely used global languages.

A Translation Model That Can Run on a Phone

One of the most notable features of the release is the extremely small deployment footprint of the 1.8B model.

Tencent uses AngelSlim 1.25‑bit quantization, which compresses the model to roughly 440 MB while preserving practical translation quality.

This allows:

• On‑device translation on common smartphone chips
• Offline translation without internet connectivity
• Faster inference compared with previous versions

According to Tencent’s reports, the quantized model also runs about 1.5× faster than the earlier Hy‑MT1.5 generation while maintaining similar deployment size.

Benchmark Performance and Comparisons

Tencent reports strong results for Hy‑MT2 across multiple translation benchmarks and real‑world scenarios.

Key claims from the technical report and release coverage include:

• 7B and 30B‑A3B models achieve state‑of‑the‑art performance among open translation models on several evaluations.
• The models perform strongly on benchmarks such as FLORES‑200, DomainMTBench, and IFMTBench, covering general, domain‑specific, and instruction‑following translation tasks.
• The 1.8B model reportedly outperforms some mainstream commercial translation APIs in aggregated benchmarks despite its much smaller size.
• In certain tests, the Hy‑MT2 family approaches the performance of leading closed systems like Gemini‑class translation models on FLORES‑200 averages.

Because most benchmark claims come from Tencent’s own report and release materials, independent third‑party evaluations are still limited. However, the reported results suggest Hy‑MT2 is particularly competitive among open translation models, especially relative to its parameter size.

Improvements Over Earlier Hunyuan Translation Models

Hy‑MT2 builds on Tencent’s earlier Hy‑MT1.5 translation models and introduces several improvements.

Major upgrades include:

Better instruction following
The models are optimized to interpret translation instructions such as formatting, terminology control, style changes, or structured translation tasks.

Stronger domain translation
Evaluations emphasize improvements in specialized fields such as finance, education, and other professional contexts.

Improved real‑world translation quality
The training pipeline includes large multilingual datasets and post‑training methods like distillation and reinforcement techniques to improve translation outputs in practical scenarios.

More flexible deployment options
The model family now spans cloud‑scale MoE models, balanced mid‑size models, and ultra‑lightweight edge deployments.

How Developers Can Access Hy‑MT2

Tencent has released the models and code publicly for developers.

Available access points include:

• GitHub repositories for code and model resources
• Hugging Face model hosting
• ModelScope distribution for the Chinese developer ecosystem

The models are designed to run across multiple hardware platforms including ARM and common server architectures.

The lightweight variant in particular targets local deployment scenarios, including edge devices and mobile hardware.

How Consumers Can Use Hy‑MT2

Tencent has also launched consumer products powered by the new models.

The first public interface is the “Tencent Hy Translation” (腾讯Hy翻译) WeChat mini‑program, which allows users to translate text and speech using the Hy‑MT2 system.

Key capabilities include:

• Voice input translation
• Custom translation style or instructions
• Online translation using larger models
• Offline translation using locally downloaded models

Tencent has also indicated that dedicated iOS and Android apps are planned, with support for local on‑device inference.

Why Hy‑MT2 Matters

Hy‑MT2 illustrates a broader trend in AI translation: moving from large, general‑purpose models toward specialized multilingual systems optimized for efficiency and deployment flexibility.

By combining:

• MoE scaling for high‑quality translation
• compact models for edge devices
• multilingual instruction following

Tencent is positioning Hy‑MT2 as both a developer‑friendly open translation stack and a practical on‑device translation engine.

If the reported benchmarks hold up under independent evaluation, Hy‑MT2 could become one of the most capable open‑source translation model families—especially for multilingual and mobile translation scenarios.