Tencent: Hy-MT2-1.8B

Tencent: Hy-MT2-1.8B is a compact translation model from Tencent, released on 20 August 2026. It sits in the text category and is built specifically for translation work rather than general chat or reasoning. The 1.8B parameter count makes it a small model by current standards, which matters if you want fast, focused translation rather than a large general-purpose assistant.

The model takes text in and returns text out, with a context window of 8192 tokens and a maximum output of 4096 tokens. Its translation coverage is broad for its size: 33 language pairs, plus five Chinese dialect and minority-language pairs. Beyond plain sentence translation, it supports several workflows, including structured translation, delimiter-based input, contextual translation, glossary-based translation, and style-guided translation. Those workflows are what separate it from a generic multilingual model, since they let you control terminology, formatting, and tone rather than just swapping words between languages.

Practical use follows directly from that feature set. Glossary-based translation fits product documentation, legal text, or any domain where terms must stay consistent across a large body of content. Delimiter-based and structured workflows suit batch jobs where you feed many segments in one request and need the output to stay aligned with the input. Contextual translation helps with dialogue, support tickets, or documents where a sentence only makes sense given the surrounding text. Style-guided translation is useful when the same content has to read differently for marketing, technical, or formal audiences. The Chinese dialect and minority-language pairs cover cases that many translation models skip entirely.

You call it through api14's OpenAI-compatible API using the model id "tencent/hy-mt2-1.8b". Because the interface follows the OpenAI chat completions shape, existing SDKs and client code work with only the model name changed. The supported request parameters are max_completion_tokens, max_tokens, stop, and temperature. Note that there is no top_p or presence penalty support here, so keep sampling control to temperature and cut generation off with stop sequences or a token cap. The 8192 token context and 4096 token output ceiling are the main practical bounds to design around.

Given the 1.8B size and the translation-specific focus, this model is a poor fit for open-ended reasoning, code generation, or long document summarisation where a larger general model would do better. It is also not the right choice if you need input beyond text, since it accepts text only. Where it does fit is high-volume translation with controlled terminology and predictable output structure, especially when latency and throughput matter more than broad world knowledge.