Tencent is pushing AI translation closer to the device itself with Hy-MT1.5-1.8B-1.25bit, an open-sourced model built to run entirely offline on smartphones. The company claims the compact model beats Google Translate while covering a broad set of languages and translation directions.
What Tencent Released
The model is called Hy-MT1.5-1.8B-1.25bit. It is designed for machine translation, but the main point is not only what it translates. The notable part is where it can run: directly on a phone, without needing an online connection.
Tencent says the model covers 33 languages, including German, English, Chinese, Japanese, French, Tibetan, and Mongolian. It also supports five dialects and 1,056 translation directions. That means the model is not limited to a small set of language pairs.
The source also says Tencent has open-sourced the model. That matters because open access can make it easier for developers, researchers and product teams to examine how the model performs and consider where offline AI translation could fit into their own tools.
Why The 440 MB Size Matters
The model’s size is central to the announcement. Tencent reduced it from 3.3 GB to 440 MB, a major drop for a translation system meant to operate on smartphones. The company achieved this through aggressive compression at 1.25 bits per parameter.
Smaller size changes the practical question around AI translation. Instead of assuming that translation must always depend on a cloud service, a compact model can be stored and run on a device. That makes offline translation more plausible for phone apps and for situations where connectivity is limited or unavailable.
The source says the 1.25-bit approach is about 25 percent smaller and 10 percent faster than earlier 1.67-bit approaches, with no quality loss. Those details point to the tradeoff Tencent is trying to solve: keep the model small enough for phones while preserving useful translation quality.
- Model size: 440 MB after compression.
- Original size: 3.3 GB before compression.
- Compression level: 1.25 bits per parameter.
- Comparison: about 25 percent smaller and 10 percent faster than earlier 1.67-bit approaches.
Performance Claims And Benchmarks
Tencent says Hy-MT1.5-1.8B-1.25bit matches commercial services and much larger models on standard benchmarks. The source specifically mentions Qwen3-32B as an example of a much larger model used for comparison.
The company also says the model has taken 30 first-place finishes in international machine translation competitions. That claim gives Tencent another way to frame the release: not just as a small model, but as one that has been evaluated against serious translation systems.
Still, the most important wording is that these are Tencent’s claims as reported in the source. The article does not provide the full benchmark tables or competition details, so the careful reading is simple: Tencent is presenting the model as competitive with established translation services and larger AI systems, while being small enough to run offline on phones.
How It Could Be Used On Android
Tencent also offers an Android demo app as an APK download. According to the source, the app translates words across any app offline. That suggests a use case beyond a standalone translation screen: translation that can work inside the phone’s broader app environment.
For users, the immediate value of offline translation is straightforward. If the model is already on the device, translation does not have to wait on a network request. It can also continue working when the phone has no connection.
For developers, the release points toward a different product design pattern. Instead of sending every translation request to a remote service, an app could rely on a local AI model for at least some translation tasks. The source does not describe integration details, so the practical scope depends on what Tencent has made available and how developers choose to use it.
A Wider Move Toward Local AI
The source places Tencent’s release in a broader trend: more AI capabilities are moving onto phones. It notes that Google is pushing in the same direction with Gemma 4, which runs locally on smartphones.
That comparison helps explain why a 440 MB translation model is significant. The race is not only about making larger models. It is also about making models that are small, fast and useful enough to run where people already work: on their own devices.
Hy-MT1.5-1.8B-1.25bit is one example of that shift. Its pitch is clear: offline AI translation, broad language coverage and a compressed footprint that Tencent says keeps quality intact. If those claims hold up in real use, smartphone translation could become less dependent on the cloud and more available wherever the device itself can operate.