Tencent’s Hy‑MT2: Multilingual AI Translation Models Built for Cloud and Mobile

All three share the same architecture goal: handling real‑world translation tasks and following translation instructions written in multiple languages.

The largest variant uses a Mixture‑of‑Experts architecture, which activates only a subset of specialized neural networks during inference. This design increases the model’s effective capacity while keeping compute costs lower than a fully dense model of similar scale.

Language Coverage

Tencent reports that Hy‑MT2 supports bidirectional translation across 33 languages.

The system also includes translation support for five Chinese ethnic languages or dialect variants, improving coverage for multilingual communication inside China’s diverse linguistic landscape.

While public summaries confirm the 33‑language capability, the complete official list of supported languages is not fully enumerated in widely available summaries of the technical report. The model family is particularly optimized for Chinese‑centric translation scenarios while also supporting widely used global languages.

A 440 MB Model That Runs on a Phone

One of the most unusual aspects of the release is the extremely small footprint of the Hy‑MT2‑1.8B model.

Tencent compresses it using AngelSlim 1.25‑bit quantization, shrinking the model to roughly 440 MB while keeping usable translation quality.

This allows several practical capabilities:

• On‑device translation on mainstream smartphone chips
• Offline translation without internet connectivity
• Faster inference compared with previous model generations

Tencent reports the quantized model runs about 1.5× faster than the earlier Hy‑MT1.5 generation while maintaining a similarly compact deployment size.

Benchmark Performance

According to Tencent’s technical report and release materials, Hy‑MT2 performs strongly across multiple translation benchmarks and real‑world evaluation tasks.

Key reported results include:

• The 7B and 30B‑A3B models achieving leading performance among open translation models on several benchmarks.
• Strong results on FLORES‑200, DomainMTBench, and IFMTBench, which evaluate general translation, domain‑specific translation, and instruction‑following translation tasks.
• The lightweight 1.8B model outperforming some mainstream commercial translation APIs in aggregated benchmark comparisons despite its smaller size.
• FLORES‑200 averages approaching the performance of leading closed systems such as Gemini‑class translation models in some evaluations.

Most benchmark claims currently come from Tencent’s own report and launch coverage, so independent third‑party validation is still limited. However, the 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.

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

Stronger domain translation
Evaluations show improved performance in specialized areas like finance, education, and professional documentation.

Higher real‑world translation quality
The training pipeline uses large multilingual datasets along with post‑training techniques such as distillation and reinforcement methods to improve practical translation accuracy.

More flexible deployment options
The Hy‑MT2 lineup now spans lightweight edge models, balanced mid‑size models, and a large MoE architecture for cloud environments.

How Developers Can Use Hy‑MT2

Tencent has released the models and associated code publicly for developers.

They are available through:

• GitHub repositories containing code and model resources
• Hugging Face model hosting
• ModelScope, a popular model platform in China’s developer ecosystem

The models are designed to run across a wide range of hardware, including ARM processors and common server architectures.

The lightweight 1.8B variant is specifically aimed at local deployment scenarios, including edge devices and mobile hardware.

How Consumers Can Try It

Tencent has also launched consumer tools built on Hy‑MT2.

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

Features include:

• Voice input translation
• Custom translation style and instructions
• High‑quality 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 on‑device translation processing.

Why Hy‑MT2 Matters

Hy‑MT2 reflects a broader trend in AI translation: moving away from general‑purpose language models toward specialized translation systems optimized for efficiency and deployment flexibility.

By combining:

• cloud‑scale MoE models for top translation quality
• compact models that run directly on phones
• multilingual instruction‑following capabilities

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

If the reported benchmark results hold up in independent testing, Hy‑MT2 could become one of the most capable open‑source translation model families—especially for multilingual and mobile translation use cases.