# Hy-MT2-30B-A3B by Tencent: Open-Weight Model
Source: https://savrn.com/models/hy-mt2-30b-a3b
Markdown alternate of the page above; the site index is https://savrn.com/llms.txt

---

## Runs On

What it takes to serve Hy-MT2-30B-A3B (30.1B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.

| Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits |
| --- | --- | --- | --- | --- | --- |
| 16-bit | 60.1 GB | 72.2 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x H100](https://savrn.com/ai-index/pricing/gpus/h100) $1.99 · [1x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $2.00 |
| 8-bit | 30.1 GB | 36.1 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x H100](https://savrn.com/ai-index/pricing/gpus/h100) $1.99 · [1x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $2.00 |
| 4-bit | 15.0 GB | 18.0 GB | 1x [MI300X](https://savrn.com/ai-index/pricing/gpus/mi300x) (192 GB) Vultr | $1.85 | [1x H100](https://savrn.com/ai-index/pricing/gpus/h100) $1.99 · [1x MI325X](https://savrn.com/ai-index/pricing/gpus/mi325x) $2.00 |

Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the [SAVRN Index](https://savrn.com/ai-index/pricing/gpus), read Oct 24, 2026.

[Hy-MT2-30B-A3B on every accelerator the SAVRN Index prices, at every precision](https://savrn.com/models/hy-mt2-30b-a3b/gpus)

## Model Card

By Tencent, published under apache-2.0, revision d3ead4dba61c.

English ｜ [中文](https://huggingface.co/tencent/Hy-MT2-30B-A3B/blob/main/README_CN.md)

[![HuggingFace](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Tencent%20Hy-ffc107?color=ffc107&logoColor=white)](https://huggingface.co/collections/tencent/hy-mt2) [![ModelScope](https://img.shields.io/badge/ModelScope-Tencent%20Hy-624aff)](https://modelscope.cn/collections/Tencent-Hunyuan/Hy-MT2)

[Official Website](https://aistudio.tencent.com/modelSquare/home/list) | [GitHub](https://github.com/Tencent-Hunyuan/Hy-MT2) | [AngelSlim](https://github.com/Tencent/AngelSlim/tree/main) | [Hy-MT2 Report](https://arxiv.org/pdf/2605.22064)

### Model Introduction

Hy-MT2 is a family of “fast-thinking” multilingual translation models designed for complex real-world scenarios. It includes three model sizes: 1.8B, 7B, and 30B-A3B (MoE), all of which support translation among 33 languages and effectively follow translation instructions in multiple languages. For on-device deployment, AngelSlim 1.25-bit extreme quantization reduces the storage requirement of the 1.8B model to only 440 MB and improves inference speed by 1.5x. Multi-dimensional evaluations show that Hy-MT2 delivers outstanding performance across general, real-world business, domain-specific, and instruction-following translation tasks. The 7B and 30B-A3B models outperform open-source models such as DeepSeek-V4-Pro and Kimi K2.6 in fast-thinking mode, while the lightweight 1.8B model also surpasses mainstream commercial APIs from providers such as Microsoft and Doubao overall.

In this release, we also open-source IFMTBench, a benchmark for evaluating translation instruction-following capabilities.

[Read the full model card (1,215 words)](https://savrn.com/models/hy-mt2-30b-a3b/card)

## Configuration

Architecture

HYV3ForCausalLM

Context length (tokens)

262,144

Layers

48

Hidden size

2,048

Feed-forward size

6,912

Attention heads

32

Key/value heads

4

Head dimension

128

Vocabulary size

120,832

Experts

128

Experts active per token

8

RoPE base

1.11588e+07

Model type

hy_v3

## Identity and Version

Repository

tencent/Hy-MT2-30B-A3B

Publisher

Tencent

Task

Translation

Modality

Text

Library

transformers

Parameters

30.1B parameters

Languages

zh, en, fr, pt, es, ja, tr, ru

Revision

d3ead4dba61c09aac60a261a96ad1df3e705febb

First published

2026-05-11

Last updated

2026-07-21

## Files and Weights

102 files, 60.1 GB in total. The weights are 51 files totalling 60.1 GB in safetensors.

Weights51 files · 60.1 GB

Configuration28 files · 1.7 MB

Tokenizer2 files · 9.7 MB

Documentation5 files · 69.8 KB

Other14 files · 4.0 MB

Repository2 files · 11.6 KB

Every file

| File | Type | Size | SHA-256 |
| --- | --- | --- | --- |
| model-00000-of-00051.safetensors | Weights | 1.2 GB | 793bf864a477 |
| model-00001-of-00051.safetensors | Weights | 1.2 GB | 9dc3d6079429 |
| model-00002-of-00051.safetensors | Weights | 1.2 GB | c83e7907b71d |
| model-00003-of-00051.safetensors | Weights | 1.2 GB | a1bae45d3b62 |
| model-00004-of-00051.safetensors | Weights | 1.2 GB | 363978d8b203 |
| model-00005-of-00051.safetensors | Weights | 1.2 GB | ace002c37643 |
| model-00006-of-00051.safetensors | Weights | 1.0 GB | b0e97f7f688e |
| model-00007-of-00051.safetensors | Weights | 1.2 GB | 1f31994e207c |
| model-00008-of-00051.safetensors | Weights | 1.2 GB | c5abdbbc0ece |
| model-00009-of-00051.safetensors | Weights | 1.2 GB | 3f3829f1cf8b |
| model-00010-of-00051.safetensors | Weights | 1.2 GB | 01048840c438 |
| model-00011-of-00051.safetensors | Weights | 1.2 GB | f748b17cd7fc |
| model-00012-of-00051.safetensors | Weights | 1.2 GB | fc9aa360c12c |
| model-00013-of-00051.safetensors | Weights | 197.1 MB | 57da846a4836 |
| model-00014-of-00051.safetensors | Weights | 1.2 GB | 9689b63963b0 |
| model-00015-of-00051.safetensors | Weights | 1.2 GB | f6599b3d7cd0 |
| model-00016-of-00051.safetensors | Weights | 1.2 GB | 067902b68230 |
| model-00017-of-00051.safetensors | Weights | 1.2 GB | fbaf386189ed |
| model-00018-of-00051.safetensors | Weights | 1.2 GB | a71fe5562184 |
| model-00019-of-00051.safetensors | Weights | 1.2 GB | 58c804ba58ac |
| model-00020-of-00051.safetensors | Weights | 1.2 GB | 9684324fa3f4 |
| model-00021-of-00051.safetensors | Weights | 1.2 GB | 46bb25f496f1 |
| model-00022-of-00051.safetensors | Weights | 1.2 GB | 9c51289c6c8e |
| model-00023-of-00051.safetensors | Weights | 1.2 GB | dc6dabec7c5b |
| model-00024-of-00051.safetensors | Weights | 1.2 GB | 36c957c06117 |
| model-00025-of-00051.safetensors | Weights | 1.2 GB | 44c31661f31f |
| model-00026-of-00051.safetensors | Weights | 1.2 GB | 89f78fba4189 |
| model-00027-of-00051.safetensors | Weights | 1.2 GB | ff2984fa0532 |
| model-00028-of-00051.safetensors | Weights | 1.2 GB | 949bb1abd62f |
| model-00029-of-00051.safetensors | Weights | 1.2 GB | c6e507472e75 |
| model-00030-of-00051.safetensors | Weights | 1.2 GB | f1117e0aff6b |
| model-00031-of-00051.safetensors | Weights | 1.2 GB | 4b78b046f61b |
| model-00032-of-00051.safetensors | Weights | 1.2 GB | 74c124ae60be |
| model-00033-of-00051.safetensors | Weights | 1.2 GB | f40965ad05fd |
| model-00034-of-00051.safetensors | Weights | 1.2 GB | 006850bd2c92 |
| model-00035-of-00051.safetensors | Weights | 1.2 GB | bdb5ac8c7f2a |
| model-00036-of-00051.safetensors | Weights | 1.2 GB | ab1be4691361 |
| model-00037-of-00051.safetensors | Weights | 1.2 GB | 7ce09cf8bb8d |
| model-00038-of-00051.safetensors | Weights | 1.2 GB | daceff3b9015 |
| model-00039-of-00051.safetensors | Weights | 1.2 GB | 7bc2f3b59fcc |
| model-00040-of-00051.safetensors | Weights | 1.2 GB | 4155e48a4dbc |
| model-00041-of-00051.safetensors | Weights | 1.2 GB | a25103d24666 |
| model-00042-of-00051.safetensors | Weights | 1.2 GB | 7634740a9204 |
| model-00043-of-00051.safetensors | Weights | 1.2 GB | 704a4eb87ed8 |
| model-00044-of-00051.safetensors | Weights | 1.2 GB | 80ffd3747d16 |
| model-00045-of-00051.safetensors | Weights | 1.2 GB | c550af73fb8c |
| model-00046-of-00051.safetensors | Weights | 1.2 GB | 2912a04d06d2 |
| model-00047-of-00051.safetensors | Weights | 1.2 GB | cff4353e83b0 |
| model-00048-of-00051.safetensors | Weights | 1.2 GB | 75e2762c1eef |
| model-00049-of-00051.safetensors | Weights | 1.1 GB | 7663117d2f33 |
| model-00050-of-00051.safetensors | Weights | 1.1 GB | cb7b37c7b79a |
| config.json | Configuration | 1.1 KB | — |
| model.safetensors.index.json | Configuration | 1.6 MB | — |
| special_tokens_map.json | Configuration | 468 B | — |
| train/deepspeed_support/ds_zero2_no_offload.json | Configuration | 528 B | — |
| train/deepspeed_support/ds_zero2_offload.json | Configuration | 676 B | — |
| train/deepspeed_support/ds_zero3_no_offload.json | Configuration | 1.1 KB | — |
| train/deepspeed_support/ds_zero3_offload.json | Configuration | 1.1 KB | — |
| train/deepspeed_support/ds_zero3_offload_no_auto.json | Configuration | 944 B | — |
| train/deepspeed_support/merge_lora_weight.py | Configuration | 2.5 KB | — |
| train/deepspeed_support/train.py | Configuration | 25.7 KB | — |
| train/deepspeed_support/train_dense.py | Configuration | 14.9 KB | — |
| train/llama_factory_support/dataset_info.json | Configuration | 974 B | — |
| train/llama_factory_support/ds_zero2_offload.json | Configuration | 892 B | — |
| train/llama_factory_support/ds_zero2_offload_lora.json | Configuration | 795 B | — |
| train/llama_factory_support/ds_zero3_offload.json | Configuration | 954 B | — |
| train/llama_factory_support/hy_dense_1_8b_full_sft.yaml | Configuration | 1.1 KB | — |
| train/llama_factory_support/hy_dense_1_8b_lora_sft.yaml | Configuration | 1.2 KB | — |
| train/llama_factory_support/hy_dense_7b_full_sft.yaml | Configuration | 1.1 KB | — |
| train/llama_factory_support/hy_dense_7b_lora_sft.yaml | Configuration | 1.2 KB | — |
| train/llama_factory_support/hy_dense_template.py | Configuration | 3.6 KB | — |
| train/llama_factory_support/hy_v3_full_sft.yaml | Configuration | 1.1 KB | — |
| train/llama_factory_support/hy_v3_lora_sft.yaml | Configuration | 1.2 KB | — |
| train/llama_factory_support/hy_v3_patches.py | Configuration | 8.8 KB | — |
| train/llama_factory_support/hy_v3_template.py | Configuration | 1.9 KB | — |
| train/llama_factory_support/train_hy_dense.py | Configuration | 3.0 KB | — |
| train/llama_factory_support/train_hy_v3.py | Configuration | 2.5 KB | — |
| train/tools/check_converted.py | Configuration | 15.6 KB | — |
| train/tools/convert_ckpt_to_outer.py | Configuration | 23.5 KB | — |
| LICENSE.txt | Documentation | 11.6 KB | — |
| README.md | Documentation | 14.8 KB | — |
| README_CN.md | Documentation | 13.8 KB | — |
| train/README.md | Documentation | 15.4 KB | — |
| train/README_CN.md | Documentation | 14.2 KB | — |
| chat_template.jinja | Other | 11.3 KB | — |
| imgs/logo-en.png | Other | 65.9 KB | — |
| imgs/logo-zh.png | Other | 69.9 KB | — |
| imgs/main_result.png | Other | 3.8 MB | b87606817fec |
| train/data/example_data.jsonl | Other | 27.8 KB | — |
| train/deepspeed_support/merge_lora_weight.sh | Other | 269 B | — |
| train/deepspeed_support/train.sh | Other | 3.5 KB | — |
| train/deepspeed_support/train_dense.sh | Other | 4.9 KB | — |
| train/deepspeed_support/train_dense_lora.sh | Other | 5.1 KB | — |
| train/deepspeed_support/train_lora.sh | Other | 3.6 KB | — |
| train/llama_factory_support/convert_zero_to_hf.sh | Other | 5.9 KB | — |
| train/llama_factory_support/train_lf.sh | Other | 3.9 KB | — |
| train/llama_factory_support/train_lf_dense.sh | Other | 4.2 KB | — |
| train/requirements.txt | Other | 159 B | — |
| .gitattributes | Repository | 1.7 KB | — |
| .ms_upload_cache | Repository | 9.9 KB | — |
| tokenizer.json | Tokenizer | 9.5 MB | — |
| tokenizer_config.json | Tokenizer | 165.6 KB | — |

## License and Download

License

apache-2.0

Access

Open weights, no gate

Download size

60.1 GB

[Download from Tencent](https://huggingface.co/tencent/Hy-MT2-30B-A3B)

Released by Tencent through its official repository on Hugging Face. [Read the license](https://www.apache.org/licenses/LICENSE-2.0).

## Built From

- Described by arXiv:2605.22064

## Memory Requirements

| Precision | Weights in memory |
| --- | --- |
| As published | 60.1 GB |
| 16-bit | 60.1 GB |
| 8-bit | 30.1 GB |
| 4-bit | 15.0 GB |

Weights only, from the published parameter count; the key-value cache and runtime add to this.

## Questions About Hy-MT2-30B-A3B

### How much GPU memory does Hy-MT2-30B-A3B need?

About 72.2 GB at 16-bit and 18 GB at 4-bit: the weights (30.1B parameters) plus a working margin. A long context needs more.

### What is the cheapest GPU to run Hy-MT2-30B-A3B on?

At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.

### Can I use Hy-MT2-30B-A3B commercially?

Yes. Hy-MT2-30B-A3B is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.

### What is Hy-MT2-30B-A3B's context length?

262,144 tokens, from the maximum position embeddings in its published configuration.

## Similar Models

Model · Translation

### [twi-translator-gemma4-31b](https://savrn.com/models/twi-translator-gemma4-31b)

[Ghana NLP Community](https://savrn.com/model-publishers/ghananlpcommunity)

An English-to-Twi (Akan) translation model fine-tuned on google/gemma-4-31B-it to translate English like an expert human translator. 1. Tool Calling: The model inspects the English source text, extracts key nominal terms, and emits a lookuptwitranslation tool call. 2. Offline Dictionary Lookup: The runtime looks up candidate Twi translations in an offline dictionary (478,822 English→Twi terms from SHOLA Gemini machine translations). 3. Linguistic Reasoning: The tool results are fed back into the model's thought channel (reasoningcontent), where the model reasons about morphology, syntax, and idioms. 4. Final Translation: The model outputs the accurate Twi translation, preserving the…

Open weights cc-by-nc-4.0 31.3B parameters 262,144 tokens

[View model](https://savrn.com/models/twi-translator-gemma4-31b)

Model · Translation

### [ChindaMT-4B](https://savrn.com/models/chindamt-4b)

[iApp Technology](https://savrn.com/model-publishers/iapp)

Paper: Reference-Grounded Data Curation for Instruction-Following Thai-English Machine Translation, AACL-IJCNLP 2026 Main Conference ChindaMT-4B is an open-weight Thai-English machine translation model fine-tuned from Qwen/Qwen3.5-4B on Grounded, a 1.97M-record dataset built by Reference-Grounded Data Curation (RGDC). It translates in both directions and follows auxiliary rules given in the prompt, such as terminology, register, length, and output format. It is one of three sizes in the ChindaMT family (4B, 2B, 0.8B). Plain translation. Same template for both directions; swap the language line and the source tag: With rules. Add a Rules: block between the language line and the source line.…

Open weights apache-2.0 5.2B parameters 262,144 tokens transformers

[View model](https://savrn.com/models/chindamt-4b)

Model · Translation

### [madlad400-3b-mt](https://savrn.com/models/madlad400-3b-mt)

[Google](https://savrn.com/model-publishers/google)

MADLAD-400-3B-MT is a multilingual machine translation model based on the T5 architecture that was trained on 1 trillion tokens covering over 450 languages using publicly available data. It is competitive with models that are significantly larger. Disclaimer: Juarez Bochi, who was not involved in this research, converted the original weights and wrote the contents of this model card based on the original paper and Flan-T5. Find below some example scripts on how to use the model: First, install the Python packages that are required: pip install transformers accelerate sentencepiece Usage with candle: We also provide a quantized model (1.65 GB vs the original 11.8 GB file): See the research…

Open weights apache-2.0 2.9B parameters transformers

[View model](https://savrn.com/models/madlad400-3b-mt)

Model · Translation

### [t5-3b](https://savrn.com/models/t5-3b)

[T5 community](https://savrn.com/model-publishers/google-t5)

The developers of the Text-To-Text Transfer Transformer (T5) write: T5-3B is the checkpoint with 3 billion parameters. The developers write in a blog post that the model: See the blog post and research paper for further details. The model is pre-trained on the Colossal Clean Crawled Corpus (C4), which was developed and released in the context of the same research paper as T5. The model was pre-trained on a on a multi-task mixture of unsupervised (1.) and supervised tasks (2.). Thereby, the following datasets were being used for (1.) and (2.): 1. Datasets used for Unsupervised denoising objective: 2. Datasets used for Supervised text-to-text language modeling objective - CoLA Warstadt et…

Open weights apache-2.0 2.9B parameters transformers

[View model](https://savrn.com/models/t5-3b)

Model · Translation

### [Hy-MT2-1.8B](https://savrn.com/models/hy-mt2-1-8b)

[Tencent](https://savrn.com/model-publishers/tencent)

English ｜ 中文 Hy-MT2 is a family of “fast-thinking” multilingual translation models designed for complex real-world scenarios. It includes three model sizes: 1.8B, 7B, and 30B-A3B (MoE), all of which support translation among 33 languages and effectively follow translation instructions in multiple languages. For on-device deployment, AngelSlim 1.25-bit extreme quantization reduces the storage requirement of the 1.8B model to only 440 MB and improves inference speed by 1.5x. Multi-dimensional evaluations show that Hy-MT2 delivers outstanding performance across general, real-world business, domain-specific, and instruction-following translation tasks. The 7B and 30B-A3B models outperform…

Open weights apache-2.0 2B parameters 262,144 tokens transformers

[View model](https://savrn.com/models/hy-mt2-1-8b)

Model · Translation

### [HY-MT1.5-1.8B-FP8](https://savrn.com/models/hy-mt1-5-1-8b-fp8)

[Tencent](https://savrn.com/model-publishers/tencent)

Hunyuan Translation Model Version 1.5 includes a 1.8B translation model, HY-MT1.5-1.8B, and a 7B translation model, HY-MT1.5-7B. Both models focus on supporting mutual translation across 33 languages and incorporating 5 ethnic and dialect variations. Among them, HY-MT1.5-7B is an upgraded version of our WMT25 championship model, optimized for explanatory translation and mixed-language scenarios, with newly added support for terminology intervention, contextual translation, and formatted translation. Despite having less than one-third the parameters of HY-MT1.5-7B, HY-MT1.5-1.8B delivers translation performance comparable to its larger counterpart, achieving both high speed and high quality.…

Open weights 1.8B parameters 262,144 tokens transformers

[View model](https://savrn.com/models/hy-mt1-5-1-8b-fp8)

## Tencent

[All models and datasets](https://savrn.com/model-publishers/tencent)

## Versions

- [d3ead4dba61c](https://savrn.com/models/hy-mt2-30b-a3b/versions/d3ead4dba61c) · current 2026-09-18

## Explore More

- [All translation models](https://savrn.com/models/tasks/translation)
- [All models under apache-2.0](https://savrn.com/models/licenses/apache-2-0)
- [Model comparisons](https://savrn.com/models/comparisons)
- [The model directory](https://savrn.com/models)
- [Open model prices by host](https://savrn.com/ai-index/pricing/open-models)

## Source

- Repository metadata, read 2026-10-21.
- [Tencent website](https://github.com/Tencent-Hunyuan)
- [Hugging Face record](https://huggingface.co/tencent/Hy-MT2-30B-A3B)
- [How the hub is built](https://savrn.com/model-hub/methodology)