---
license: apache-2.0
base_model: tencent/Hy-MT2-1.8B-GGUF
pipeline_tag: translation
library_name: openasr
tags:
  - translation
  - simultaneous-translation
  - openasr
  - oasr
---

<div align="center">

# Hy-MT2 1.8B · OpenASR

**Hy-MT2 1.8B zh→en translation for OpenASR realtime subtitles — clause-by-clause, fully on-device**

[![License](https://img.shields.io/badge/license-Apache--2.0-2563eb.svg)](https://huggingface.co/tencent/Hy-MT2-1.8B-GGUF/blob/1cd5208700acedef4ef93019b6cfc148b8522d45/LICENSE.txt)
[![Format](https://img.shields.io/badge/format-.oasr-7c3aed.svg)](https://github.com/QuintinShaw/openasr)
[![Runtime](https://img.shields.io/badge/runtime-OpenASR-111827.svg)](https://openasr.org)
[![Base model](https://img.shields.io/badge/base-Hy--MT2--1.8B--GGUF-f59e0b.svg)](https://huggingface.co/tencent/Hy-MT2-1.8B-GGUF)

Realtime translation pack for the **[OpenASR](https://github.com/QuintinShaw/openasr)** runtime —
pure-Rust inference, **no Python at inference time**.

</div>

---

## ✨ Highlights

- 🌐 **Realtime zh→en subtitle translation** — drives OpenASR's experimental clause-retranslation lane next to live Chinese ASR
- ⚡ **Hot-session clause latency in the hundreds of milliseconds** — prefix KV-cache reuse keeps warm per-clause retranslation fast on Apple Silicon CPU
- 📦 **Byte-identical upstream weights** — the Q4_K_M GGUF tensor data is repackaged unmodified at a pinned revision; sha256 and provenance are embedded in the pack
- 🧾 **License + NOTICE embedded** — the upstream Apache-2.0 LICENSE.txt and the OpenASR modification NOTICE travel inside the pack metadata
- 🎯 **Parity-gated packaging** — greedy decode token IDs match llama.cpp on the same prompt before release
- 🦀 **Native in OpenASR** — `.oasr` packs run with no Python at inference, engineered for peak performance on CPU & GPU

## 🚀 Quickstart

```bash
# 1. Install the OpenASR CLI  ·  https://openasr.org
# 2. Pull the pack
openasr pull hymt2-1.8b:q4km
```

Realtime translation is an experimental capability: with this pack installed,
realtime sessions can enable clause-by-clause zh→en translation alongside live
Chinese ASR transcripts.

## 📦 Pack

| Quant | File (`.oasr`) | Size |
|:------|:---------------|-----:|
| Q4_K_M | `hymt2-1.8b-q4_k_m.oasr` | 1.13 GB |

<sub>Single **q4_k_m** build repackaged from the upstream GGUF release. The
tensor data is preserved byte-for-byte — no requantization, no weight changes.</sub>

## 🧠 About Hy-MT2 1.8B

Hy-MT2 1.8B is Tencent's compact many-to-many translation model from the Hunyuan-MT family.
OpenASR packages the upstream Q4_K_M GGUF release as a local `.oasr` translation pack and uses it
for the experimental simultaneous-translation MVP: finalized and in-progress Chinese clauses from
realtime ASR are retranslated clause-by-clause into English, with provisional translations
superseded as the source clause stabilizes. The MVP contract is zh→en; the pack runs fully
on-device through OpenASR's pure-Rust GGML runtime with no Python at inference time.

## ⚙️ How this pack was made

Repackaged from [tencent/Hy-MT2-1.8B-GGUF](https://huggingface.co/tencent/Hy-MT2-1.8B-GGUF) at a pinned revision with
the OpenASR importer:

```bash
openasr model-pack import hymt2-gguf <src>.gguf <out>.oasr \
  --package-id hymt2-1.8b \
  --license-file LICENSE.txt --notice-file NOTICE.openasr.txt
```

The `.oasr` container is GGUF-backed. The importer verifies the pinned upstream
sha256 and splices provenance, license, and notice metadata into the GGUF KV
section; the model weights are byte-identical to the upstream GGUF artifact.

## ⚖️ License

This pack **inherits the upstream model's license: Apache-2.0**
([source](https://huggingface.co/tencent/Hy-MT2-1.8B-GGUF/blob/1cd5208700acedef4ef93019b6cfc148b8522d45/LICENSE.txt)). The upstream `LICENSE.txt` and the OpenASR
modification `NOTICE.openasr.txt` are embedded in the pack metadata; the only
modification is packaging metadata, not weights.

## 🙏 Acknowledgements

This pack redistributes **tencent/Hy-MT2-1.8B-GGUF** in OpenASR's `.oasr` runtime format. Credit
for the model architecture, training, and original weights belongs to the upstream Tencent Hunyuan
authors. The upstream model is licensed under **Apache-2.0**; OpenASR packaging retains that
license and attribution. The model weights are byte-identical to the upstream GGUF artifact —
the only modification is packaging metadata, recorded in the embedded `NOTICE.openasr.txt`.

## 🔗 Links

- 🦀 **OpenASR** — <https://github.com/QuintinShaw/openasr>
- 🌐 **Website** — <https://openasr.org>
- 🤗 **Upstream model** — [tencent/Hy-MT2-1.8B-GGUF](https://huggingface.co/tencent/Hy-MT2-1.8B-GGUF)