Tencent Hy-MT2-30B-A3B

Hy-MT2-30B-A3B is the largest of three sizes in Tencent's second-generation machine translation family, published on 21 May 2026 together with the 1.8B and 7B models. Unlike the general-purpose chat models that dominate this catalogue, it is a single-purpose system: it translates, and Tencent describes the whole line as "fast-thinking", meaning it answers directly instead of producing a chain of reasoning first. The model covers 33 languages, Polish among them, and the interesting part is what the instruction handling adds on top. Tencent documents prompt patterns for supplying a glossary of required term translations, for enforcing a named style, for carrying user preferences across a job, for preserving delimiters exactly, and for translating only the human-readable values inside JSON or XML while leaving keys, tags and placeholders untouched. That last case is the difference between a translation model and a translation tool for software teams. Alongside the weights Tencent published IFMTBench, its own open benchmark for measuring exactly this instruction-following behaviour. On its own charts Tencent puts the 7B and 30B-A3B ahead of general open models such as DeepSeek-V4-Pro and Kimi K2.6 when those are run in fast-thinking mode. As always with a vendor benchmark, the comparison is chosen by the party being measured; the independent check here is that the weights are downloadable and the benchmark is published. For a reader in the European Union the licence matters as much as the scores. We read the LICENSE.txt file in the repository: it is the plain Apache 2.0 text with no added territorial clause. This is not a given at Tencent — the company's HunyuanOCR models, also in this catalogue, ship under a community licence that expressly does not apply in the EU, the UK and South Korea. The translation family carries no such exclusion.

▸Videos

No videos yet.

▸Reviews