1. Translation Across 33 Languages
Hy-MT2-1.8B supports translation across 33 languages, along with several Chinese minority-language and dialect translation scenarios, including languages such as Uyghur and Tibetan.
Most major languages are covered, so the model is not limited to the usual Chinese-English or English-Japanese pairs.
Many users will never need the full language list, but the broader coverage can matter for multilingual apps or products aimed at specific regions.
2. Translation Instructions and Terminology Control
This is more useful than a small improvement on a benchmark score.
You can tell the model in advance how you want a translation handled. For example:
- Use a fixed translation for a specific term
- Make the result more conversational or more formal
- Leave certain words untranslated
- Preserve a particular output format
- Follow a preferred terminology convention
Take a product document containing the words “workspace,” “deployment,” and “endpoint.”
A basic translation system may translate the same term differently from one paragraph to another. With Hy-MT2, you can specify that “workspace” should always use the same translated term, “deployment” should follow your preferred terminology, and “endpoint” should remain in English.
The result may not make every sentence perfect, but at least the terminology stays consistent.
That difference is easy to miss when translating a short message. It becomes much more important in technical documentation, product manuals, or business material, where inconsistent terminology creates a lot of cleanup work.
Tencent has also released IFMTBench, a benchmark designed to test how well translation models follow these kinds of instructions.
3. A 440MB Model That Can Actually Run on a Phone
The quantized Hy-MT2-1.8B model is roughly 440MB.
That is small enough to make direct integration into a mobile app realistic, with inference happening on the device instead of in the cloud.
On Apple’s A15 chip, Tencent reports inference performance around 1.5 times faster than the previous generation.
For short translations, the important part is that you do not have to wait for a network round trip. The model can keep working on a plane, in a subway tunnel, or anywhere else with poor connectivity.
There is a trade-off, though.
The smallest versions rely on aggressive quantization. A 1.25-bit quantized model saves a lot of space, but compression at that level can reduce translation quality in some cases.
So the 440MB figure is impressive, but it should not be read as “440MB with zero compromise.”
4. Improvements Over Hy-MT1.5
On Tencent’s real-world translation benchmark, Hy-MT2-1.8B increased its GEMBA score from 88.82 to 91.08 compared with the previous generation.
Benchmark numbers never tell the whole story, but the increase points to better translation accuracy and more natural output, with fewer awkward machine-translated phrases.
The size reduction may be even more important for mobile use.
The previous generation required more than 1GB of storage, while the quantized Hy-MT2-1.8B comes in at around 440MB.
On a server, saving a few hundred megabytes is not particularly exciting. Inside a mobile app, it affects download size, storage requirements, memory pressure, and the range of devices you can reasonably support.
That is where the generational upgrade becomes much easier to appreciate.
Comments (0)