The middle size of Tencent's open translation family: eight billion dense parameters across 33 languages, meant for a single server card rather than a phone or a cluster.
Hy-MT2-7B sits between the 1.8B model built for phones and the 30B mixture-of-experts built for servers, and it was published with both on 21 May 2026. It is a dense model of 8.03 billion parameters covering the same 33 languages and the same instruction-driven translation behaviour as the rest of the family: enforced glossaries, named styles, exact delimiter preservation, and structured-data translation that changes the human-readable values in JSON or XML while leaving keys, tags and placeholders alone.
Tencent groups it with the 30B-A3B model in its own comparison, claiming both beat general open models such as DeepSeek-V4-Pro and Kimi K2.6 when those are run in fast-thinking mode — that is, answering directly rather than reasoning first. It is a vendor chart, and the choice of comparison belongs to the vendor; what is independently checkable is that the weights and the company's IFMTBench instruction-following benchmark are both published.
The practical case for this size is that it is the largest of the three that still runs comfortably as a dense model on a single accelerator, without the serving complexity of expert routing. Over the 30 days to 31 August 2026 the llama.cpp build drew 29,691 downloads against 12,052 for the base weights.
The repository licence is the plain Apache 2.0 text, checked in the LICENSE.txt file — with none of the territorial exclusion that Tencent applies to its HunyuanOCR models against the EU, the UK and South Korea.