Hy-MT2: Tencent's 440 MB offline translation models, all three sizes open
Tencent Hunyuan's Hy-MT2 translation models span 1.8B to 30B-A3B across 33 languages; the 1.8B fits in 440 MB after 1.25-bit quantization and runs on phones.
Project facts
GitHub Ecosystem- License
- Apache-2.0
- Language
- Python
- Stars
- 1,758
- Data checked
- 2026-11-15
Snapshot figures reflect the check date and may change over time.
Translation is one of the few model tasks where you can judge the result yourself, which is why open translation models deserve a closer look than leaderboard numbers: run your own text through them and you will know. Hy-MT2 is Tencent Hunyuan’s “fast-thinking” translation family — 1.8B, 7B and 30B-A3B (MoE), all translating among 33 languages and following translation instructions given in multiple languages. The repository is Apache-2.0 (LICENSE.txt is the standard Apache 2.0 text; GitHub labels it NOASSERTION because of the file name). GitHub’s API reported 1,758 stars on November 15, 2026.
Core features
- Three sizes for three compute budgets: 1.8B targets on-device and consumer hardware; 7B is the general workhorse; 30B-A3B is a MoE (about 3B active) for server-grade quality.
- On-device quantization: the README says AngelSlim 1.25-bit extreme quantization brings the 1.8B model down to 440 MB of storage with 1.5x faster inference; FP8 and GGUF (llama.cpp) variants are published too, including 2-bit and 1.25-bit builds of the 1.8B.
- Instruction-following translation: the models translate according to source language, target language and domain terms in the prompt. The README requires full language names (Chinese names in Chinese prompts), and the models ship without a default system prompt.
- Server deployment: the README gives a vLLM path (
vllm serve tencent/Hy-MT2-30B-A3B --tensor-parallel-size 1) and an SGLang launch command, with recommended inference parameters for the 1.8B/7B and the 30B-A3B. - A companion benchmark: IFMTBench, the instruction-following translation benchmark, was open-sourced alongside the models under its own license file. The line has three generations — Hunyuan-MT-Chimera-7B (September 2025), HY-MT1.5 (December 2025) and this release (May 2026).
Typical use cases
- Offline and privacy-sensitive translation: text never leaves the device — travel, field work, or anything that cannot go online; the GGUF variants run in the llama.cpp ecosystem directly.
- Domain-specific batch translation: glossaries and domain instructions go into the prompt; the 7B or 30B-A3B can power reproducible translation pipelines.
- A self-hosted alternative to translation APIs: the README claims the 1.8B surpasses mainstream commercial APIs from providers such as Microsoft and Doubao overall — a vendor claim, so test on your own corpus before adopting. If the job is translating whole PDF documents rather than raw text, pair these weights with a document pipeline such as babeldoc, which handles parsing, layout preservation and batch translation on top of a model.
Quick start
Weights are on the Hugging Face collection and mirrored on ModelScope. The fastest path is the 1.8B: download Hy-MT2-1.8B or a GGUF variant, load it locally, and send prompts in the README’s format — full source and target language names plus the text to translate. Recommended sampling settings are in the README, and remember there is no default system prompt. On a server, vLLM or SGLang starts the 30B-A3B with one command.
Summary
Hy-MT2 pushes offline translation down to 440 MB of storage and phone-grade hardware, and the three sizes give you a ladder from device to server. Two facts belong in any evaluation: the benchmark numbers (beating DeepSeek-V4-Pro, Kimi K2.6 and mainstream commercial APIs) are all vendor-reported with no independent replication, and the repository was released in May 2026 with no pushes since late August — treat it as usable open weights rather than a fast-moving product. GitHub’s API reported Python, Apache-2.0, 1,758 stars, an unarchived repository, last pushed August 26, 2026, on November 15, 2026.
This entry is based on the README and model cards; we did not load the models or run translations. The quantization size and speed figures come from AngelSlim and the README — on-device results depend on your hardware. For a production pipeline, treat these weights as the sentence-level engine and let a document layer own layout and batching.