---
language:
- zh
- en
- ja
- ko
- ar
- de
- fr
- es
- ru
- bn
- cs
- et
- fa
- he
- id
- kn
- ro
- ta
- te
- tl
- tr
- uk
library_name: transformers
license: apache-2.0
base_model: openbmb/MiniCPM5-2B
base_model_relation: finetune
pipeline_tag: translation
tags:
- translation
- minicpm
- minicpm5
- gguf
- on-device
---
🤗 Hugging Face |
💾 GGUF
# MiniTranslator-2B
## Model Introduction
MiniTranslator-2B is a 2B multilingual translation model built on
[MiniCPM5-2B](https://huggingface.co/openbmb/MiniCPM5-2B). It supports translation
among Chinese, English, and 20 additional languages, and follows translation
instructions in both Chinese and English. The model is designed for general
translation workloads as well as on-device and edge deployment.
MiniTranslator-2B is released in BF16 and in four GGUF variants (F16 / Q8_0 / Q6_K /
Q4_K_M), so the same model can run with transformers, vLLM, SGLang, or llama.cpp.
## Results
All translation results use greedy decoding (`temperature=0`) and XCOMET (higher is
better). Instruction-following results use the benchmark's rule-based scoring.
### Translation

MiniTranslator-2B is competitive with Hy-MT2-1.8B on FLORES-200, NTREX-128, and
WMT24++, and leads on HardMTBench by 3.1 points. WMT25 long-form and structured
content remains the largest gap to the 7B and 30B-A3B models.
### Instruction Following

## Model Links
| Model | Description | Download |
|---|---|---|
| MiniTranslator-2B | BF16 weights | [🤗 Model](https://huggingface.co/Caldalis/MiniTranslator-2B) |
| MiniTranslator-2B-GGUF | F16 / Q8_0 / Q6_K / Q4_K_M | [🤗 Model](https://huggingface.co/Caldalis/MiniTranslator-2B-GGUF) |
## Translation Task Instruction Examples (Chinese-English Comparison)
*Note: both `source_lang` and `target_lang` should use full language names. Use
Chinese names in Chinese prompts and English names in English prompts.*
| Type | Chinese prompt | English prompt |
|---|---|---|
| **Default Translation** | 将以下文本翻译为 `{target_lang}`,只输出译文:
`{source_text}` | Translate the following text into `{target_lang}`. Output only the translation:
`{source_text}` |
| **Terminology** | *参考下面的翻译:*
`{text}` 翻译成 `{text}`
`{text}` 翻译成 `{text}`
将以下文本翻译为 `{target_lang}`,只输出译文:
`{source_text}` | *Reference the following translations:*
`{text}` translates to `{text}`
`{text}` translates to `{text}`
Translate the following text into `{target_lang}`. Output only the translation:
`{source_text}` |
| **Style** | 请将以下文本翻译为 `{target_lang}`。
翻译风格要严格符合【**`{target_style}`**】
`{source_text}` | Please translate the following text into `{target_lang}`. The translation style must strictly conform to [**`{target_style}`**]:
`{source_text}` |
| **Personalization** | *【待翻译文本】*
`{source_text}`
*【翻译任务】*
1、**`{user_preferences}`**
2、**`{user_preferences}`**
3、……
4、将【待翻译文本】翻译为 `{target_lang}`。 | *[Source Text]*
`{source_text}`
*[Translation Tasks]*
1. **`{user_preferences}`**
2. **`{user_preferences}`**
3. ...
4. Translate the [Source Text] into `{target_lang}`. |
| **Delimiters** | 请将以下文本准确翻译为 `{target_lang}`。
你必须在译文中**保留等量的分隔符,绝对不可遗漏、转义或翻译该符号,并注意分隔符的位置**。
`{source_text}` | Please accurately translate the following text into `{target_lang}`.
You must **retain the exact same number of delimiters in the translation. Strictly do not omit, escape, or translate these symbols, and pay close attention to their placement**.
`{source_text}` |
| **Structured Data** | *# 任务目标*
将下方 `{source_text}` 中的 `{format_type}` 格式数据翻译为 `{target_lang}`。
*# 严格约束*
1. **结构锁定**:绝对保持原有的 `{format_type}` 数据结构、缩进和层级完全不变。
2. **选择性翻译**:仅翻译面向用户展示的可见文本内容。
3. **禁止修改**:**严禁**翻译或更改任何代码标签、键名(Key)、变量占位符(如 `{{var}}`、`${var}`、`%s`、`%d` 等)或代码属性。
*# 数据输入*
`{source_text}` | *### Task*
Translate the user-facing text within the following `{format_type}` data into `{target_lang}`.
*### Strict Rules*
1. **Structure Preservation:** You MUST preserve the original `{format_type}` data structure, nesting, hierarchy, and indentation exactly as they are.
2. **Selective Translation:** Translate ONLY the visible, user-facing text content/values.
3. **Strict Non-Translation:** NEVER translate or alter code tags, keys, properties, object names, or variable placeholders. Leave them exactly in their original English/code form.
*### Source Data*
`{source_text}` |
| **Background Information** | *【背景信息】*
`{background_text}`
请结合背景信息将以下文本翻译为 `{target_lang}`。
*【待翻译文本】*
`{source_text}` | *[Background Information]*
`{background_text}`
Please translate the following text into `{target_lang}`, taking the provided background information into consideration.
*[Source Text]*
`{source_text}` |
---
## Inference and Deployment
Recommended decoding parameters. The model does not have a default system prompt.
```json
{
"temperature": 0,
"top_p": 0.95,
"max_tokens": 4096
}
```
Greedy decoding (`temperature=0`) is recommended for translation. If sampling is
needed, use a low temperature (0.2–0.3) and `top_p=0.9`.
Thinking mode is supported by the chat template but should be disabled for
translation: pass `enable_thinking=False` in transformers, `--reasoning off` in
llama.cpp, or `chat_template_kwargs: {"enable_thinking": false}` in an
OpenAI-compatible request.
### transformers
`transformers>=5.6.0`
```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_path = "Caldalis/MiniTranslator-2B"
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(
model_path,
dtype=torch.bfloat16,
device_map="auto",
)
model.eval()
prompt = "将以下文本翻译成英语,只输出译文:\n\n今天天气真好。"
messages = [{"role": "user", "content": prompt}]
inputs = tokenizer.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
enable_thinking=False,
return_tensors="pt",
).to(model.device)
with torch.no_grad():
outputs = model.generate(**inputs, max_new_tokens=4096, do_sample=False)
response = tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True)
print(response)
```
### vLLM
```bash
vllm serve Caldalis/MiniTranslator-2B --tensor-parallel-size 1
```
### SGLang
```bash
python3 -m sglang.launch_server --model Caldalis/MiniTranslator-2B --tp 1
```
### llama.cpp
```bash
llama-server -m MiniTranslator-2B-Q4_K_M.gguf -c 4096 --reasoning off --port 8080
```
```bash
curl http://127.0.0.1:8080/v1/chat/completions \
-H 'Content-Type: application/json' \
-d '{
"messages": [{"role": "user", "content": "将以下文本翻译成英语,只输出译文:\n\n今天天气真好。"}],
"temperature": 0,
"max_tokens": 4096
}'
```
On llama.cpp builds without `--reasoning`, omit the server flag and add
`"chat_template_kwargs": {"enable_thinking": false}` to the request body.
## Quantization
| File | Quantization | Size (bytes) | SHA-256 |
|---|---|---:|---|
| `MiniTranslator-2B-F16.gguf` | F16 | 5,039,006,496 | `236b69f8d40370ac390ff6f48283349d66318d654094a643d3641f148aa8e958` |
| `MiniTranslator-2B-Q8_0.gguf` | Q8_0 | 2,679,710,496 | `5999502dd1e520b18fcfa5bade61b64b7b119fa991d16f15b816f4f3b46d1997` |
| `MiniTranslator-2B-Q6_K.gguf` | Q6_K | 2,070,225,696 | `df7895713c691700c36b1db458a8998cb2ac8cd60e8cff1cb8976a3fe500dd93` |
| `MiniTranslator-2B-Q4_K_M.gguf` | Q4_K_M | 1,561,318,176 | `e7ebfd5120803c59163ed0cb44725c1e43e97be8434b031474c36ef32a867114` |
Weights are converted with `llama.cpp/convert_hf_to_gguf.py` and quantized with
`llama-quantize`. See the GGUF model repository for conversion and usage details.
## Supported Languages
| Language | Abbr. | Chinese Name |
|---|---|---|
| Chinese | zh | 中文 |
| English | en | 英语 |
| Japanese | ja | 日语 |
| Korean | ko | 韩语 |
| Arabic | ar | 阿拉伯语 |
| German | de | 德语 |
| French | fr | 法语 |
| Spanish | es | 西班牙语 |
| Russian | ru | 俄语 |
| Bengali | bn | 孟加拉语 |
| Czech | cs | 捷克语 |
| Estonian | et | 爱沙尼亚语 |
| Persian | fa | 波斯语 |
| Hebrew | he | 希伯来语 |
| Indonesian | id | 印尼语 |
| Kannada | kn | 卡纳达语 |
| Romanian | ro | 罗马尼亚语 |
| Tamil | ta | 泰米尔语 |
| Telugu | te | 泰卢固语 |
| Filipino | tl | 菲律宾语 |
| Turkish | tr | 土耳其语 |
| Ukrainian | uk | 乌克兰语 |
The base model supports additional languages beyond this list. Quality on
low-resource directions varies and should be validated on in-domain data.
## License
Released under the **Apache License 2.0**. This model is a derivative of
[openbmb/MiniCPM5-2B](https://huggingface.co/openbmb/MiniCPM5-2B); see `NOTICE` for
attribution. "MiniCPM" and any OpenBMB marks remain the property of
their owners.