LifeHubber. Resources Use AI on your terms.
LifeHubber.Resource
Opening…

AI Resources

Hy-MT2

GitHub stars: 1.8K GitHub forks: 188 Last pushed August 26, 2026: Pushed 2mo ago
Stats from GitHub

Hy-MT2 is a Tencent-Hunyuan multilingual translation model family for complex real-world translation scenarios.

Tencent-Hunyuan presents Hy-MT2 with 1.8B, 7B, and 30B-A3B variants, support for 33 languages, an AngelSlim 1.25-bit on-device option, IFMTBench, training materials, and several local and server deployment paths.

What it is

A translation-focused model family

Hy-MT2 is built around machine translation rather than general chat, with multiple model sizes and examples for default translation, terminology handling, style control, personalization, delimiters, and structured data.

Why it stands out

Broad release with local options

The release includes 1.8B, 7B, and 30B-A3B variants plus an AngelSlim 1.25-bit 1.8B option for readers comparing on-device and server translation workflows.

Availability

Models, benchmark, report, and training notes

The repository points to Hugging Face and ModelScope model pages, IFMTBench, a report PDF, translation instruction examples, deployment notes, and training documentation.

Why it matters

What makes it useful

Translation instructions can specify terminology and style, rather than leaving every phrasing choice to the model. Hy-MT2's examples include preferred source-term translations and a target style, giving a builder a way to express a glossary or tone requirement alongside the passage.

Notable points

What stands out

Tencent says AngelSlim compresses the 1.8B model to about 440 MB and reports a speed improvement. Treat those figures and the wider model comparisons as project results until they are tested on the reader's own hardware and languages.

Before using

What to review

Choose the model size and format for the intended runtime; the 1.25-bit 1.8B option and the larger server models have different deployment requirements. The selected 1.25-bit GGUF path requires an STQ-compatible llama.cpp build. Check that the build contains the required kernel; the project's linked STQ pull request is still open, so its existence does not establish support in a stock release.

The model card specifies Transformers 5.6.0 or later and uses trust_remote_code in its example. Hugging Face's security guidance calls for pinning a reviewed revision and enabling remote code only for repositories the operator has inspected.

Tencent-reported compressed size, speed and translation comparisons depend on the stated hardware, languages and evaluation setup; they do not establish results for every task.

Structured output and terminology instructions describe desired behavior. Check translations and preserved keys and substitution tokens before using generated data in an application.

Check the original listing's documentation for current requirements, limitations, licensing and any relevant privacy or security considerations.

Reader fit

Who may find it relevant

Readers comparing multilingual translation models and instruction-following translation behavior.

Builders exploring local, quantized, or serving-oriented translation workflows.

Less relevant for readers looking mainly for a general assistant, coding agent, or consumer chat product.

Editorial note

Why LifeHubber lists it

The structured-data example asks Hy-MT2 to translate user-facing values while keeping keys, code tags, substitution tokens and nesting unchanged. That is a different task from translating an entire block as prose: it is meant for content such as interface strings inside a data file. The prompt states the required preservation; it does not guarantee every generated file satisfies it.

Source links

Source materials

Reader note

About our listings

LifeHubber’s listings help you explore what’s available in AI and are not endorsements. We don’t independently verify every detail or guarantee that each listing is accurate, safe, actively maintained, or right for your needs.

What to explore next

Compare another small-device translation path.

If the next decision is a low-bit translation artifact and its runtime, compare the earlier Hy-MT1.5 path. Its published weight formats and custom-kernel instructions continue that setup question.

Advertisements

Advertisements

For project maintainers

Listed here? You can use the badge.

If you maintain a project with a current LifeHubber listing, you may add the optional “Listed on LifeHubber AI Resources” badge to its README, docs, or website. No introduction or permission request is needed.

MORE ON LIFEHUBBER

Use AI on your own terms.

Explore more AI projects, stories, guides, tools and community insights across LifeHubber.

See what’s moving