Hy-MT2-7B

Hy-MT2-7B is part of Tencent's Hy family of multilingual translation models, which is also offered in 1.8B and 30B-A3B MoE variants. The 7B release uses the Hunyuan-dense architecture and ships as roughly 7.5 billion parameters, with the community Ollama artifact weighing about 4.6GB under Q4_K_M quantization. Independent directories attribute the release to Tencent and link it to the official tencent/Hy-MT2-7B repository on Hugging Face under an Apache-2.0 license, making the weights openly available for self-hosted experimentation and downstream fine-tuning.

The model is purpose-built for real-world translation across 33 languages and is trained to follow translation instructions directly, rather than simply completing text in a prompted target language. This instruction-following orientation lets users specify tone, terminology constraints, or domain context and have the model adapt its output accordingly, which is well suited to localization workflows, multilingual customer support content, and cross-language documentation pipelines. A long context window reported at 256K tokens further enables translation of lengthy documents, code-heavy technical material, or full chapter-level passages in a single request, distinguishing it from shorter-context general-purpose chat models.