1. Domain-Specific Translation
The 30B-A3B model has been optimized for eight professional domains, including finance, law, politics, and education.
On DomainMTBench, it scored 97.08 in finance and 89.15 in law using XCOMET.
Professional translation usually breaks in small ways before it breaks in obvious ones.
A technical term is translated one way on page three and a different way on page nine. A legal sentence is technically translated word for word, but the logical structure gets muddled. A financial term is correct in isolation but inconsistent across the full document.
Those are the kinds of problems that matter in real work.
That is also where a larger, more specialized model starts to earn its keep.
2. Better Instruction Following
This is one of the clearest improvements over Hy-MT1.5.
You can give the model a rule like:
“Keep the translation concise and limit each sentence to 15 words.”
You can also provide a glossary and force certain terms to use a fixed translation.
Real projects usually come with several rules at once.
Keep the formatting. Preserve punctuation. Do not translate field names. Keep product names unchanged. Use the same tone throughout.
None of those requirements is difficult by itself. The problem starts when they are all combined.
Hy-MT2 was built to handle that kind of instruction stack more reliably. Tencent also released IFMTBench to measure this capability, and the 30B-A3B model scored 84.69 on the complex-instruction test.
For production translation, I find this kind of result more useful than another generic “fluency” score. It tells you whether the model can actually stay inside the rules you give it.
3. Performance on Real-World Translation Tasks
Hy-MT2-30B-A3B scored 89.87 on WildMTBench using GEMBA, ahead of Gemini 3.1 Pro in that benchmark.
WildMTBench is designed to look more like real translation work, where the input is less clean and predictable than standard test data.
That matters.
Actual business documents rarely look like benchmark examples. Technical files mix abbreviations, table fields, half-translated phrases, and inconsistent formatting. Commercial documents contain industry jargon. Contracts are full of nested clauses and long sentences.
A model that performs well on polished test sentences can still fall apart on that kind of material.
The WildMTBench result suggests that 30B-A3B is not just good at controlled benchmark conditions.
4. What Changed From Hy-MT1.5
The most noticeable upgrade is instruction following.
Hy-MT1.5 was more likely to miss a rule when several translation requirements were combined. Hy-MT2 is more consistent with terminology, formatting, length limits, and style constraints.
Minority-language translation also improved.
In tests covering languages such as Tibetan, Uyghur, and Mongolian, the 30B-A3B model reached an XCOMET score of 62.44, higher than Gemini 3.1 Pro in the same comparison.
The Hy-MT2 family also pushes quantization quite aggressively.
That matters less for the 30B-A3B model itself, which is better suited to server or cloud deployment. But the same work carries down to the smaller models. The 1.8B version can be compressed to around 440MB and run locally on a phone.
The product strategy is easy to see: the larger model focuses on professional translation quality, while the smaller one is pushed toward on-device use.
Comments (0)