Qingqiu-MT-9B

Model Introduction

Qingqiu-MT-9B is an instruction-following multilingual translation model from Kingsoft Office WPS-Qingqiu Team, released for the constrained track of the WMT26 General Machine Translation Shared Task.

It is based on Qwen/Qwen3.5-9B and primarily supports mutual translation among 20 languages, covering 22 of the 23 official WMT26 language pairs (excluding English→Northern Sámi). Unlike conventional MT settings that only provide source text and a target language, each sample also includes a natural-language translation instruction. The model is expected to follow that instruction and output only the translation.

License: Apache 2.0. For system details, see our WMT26 system paper (Qingqiu-MT-9B: An Instruction-Following Multilingual Translation Model).

System Overview

Figure: Qingqiu-MT-9B system overview — source-grounded instruction-translation data construction, two-stage SFT by data quality, and GRPO reinforcement learning.

Evaluation Results

We report development results on IF-WMT25 (main set; 5,239 samples covering 22 WMT26 directions) and IFMTBench multi-constraint subset (2,838 samples). Evaluation uses an LLM-as-a-judge protocol (GLM-5): for each source text and instruction, anonymous system outputs are scored jointly on Accuracy / Terminology / Format / Fluency / Style (0–100), and the overall score is a weighted sum with weights (0.40, 0.20, 0.15, 0.15, 0.10). Tables report pairwise Win / Lose / Gap (%) of Qingqiu-MT-9B vs each baseline (Gap = Win − Lose; ties counted in the denominator). Overall is sample-level aggregation over the full set.

Baselines: Qwen3.5-9B, HY-MT2-7B, Gemma4-12B, DeepSeek-v4-flash.

Overall summary

IF-WMT25

Baseline	Win	Lose	Gap
Qwen3.5-9B	83.2	13.7	+69.5
HY-MT2-7B	52.4	40.8	+11.6
Gemma4-12B	52.3	38.3	+14.0
DeepSeek-v4-flash	38.7	52.4	-13.7

IFMTBench (multi-constraint)

Baseline	Win	Lose	Gap
Qwen3.5-9B	52.5	26.6	+25.9
HY-MT2-7B	45.9	28.5	+17.4
Gemma4-12B	31.9	30.8	+1.1
DeepSeek-v4-flash	28.3	38.8	-10.5

IF-WMT25 (by direction)

Direction	Qwen3.5-9B W/L/G	HY-MT2-7B W/L/G	Gemma4-12B W/L/G	DeepSeek-v4-flash W/L/G
cs→de	81.1 / 18.9 / +62.2	45.2 / 52.4 / -7.2	44.8 / 48.0 / -3.2	29.9 / 63.8 / -33.9
cs→uk	84.1 / 15.1 / +69.0	43.9 / 52.0 / -8.1	53.4 / 42.5 / +10.9	29.9 / 63.8 / -33.9
cs→vi	80.3 / 17.3 / +63.0	61.6 / 36.8 / +24.8	46.6 / 40.6 / +6.0	46.5 / 48.0 / -1.5
en→arz	93.7 / 6.3 / +87.4	78.7 / 18.1 / +60.6	35.6 / 54.9 / -19.3	44.9 / 52.8 / -7.9
en→be	86.7 / 12.5 / +74.2	--	82.8 / 12.8 / +70.0	29.5 / 62.3 / -32.8
en→cs	90.5 / 7.9 / +82.6	39.1 / 57.0 / -17.9	63.5 / 29.4 / +34.1	29.4 / 65.1 / -35.7
en→de	86.1 / 12.0 / +74.1	56.1 / 43.0 / +13.1	34.6 / 55.7 / -21.1	41.7 / 51.9 / -10.2
en→et	88.0 / 10.4 / +77.6	--	81.1 / 16.9 / +64.2	27.2 / 68.8 / -41.6
en→hy	96.1 / 3.9 / +92.2	--	76.4 / 22.4 / +54.0	41.1 / 57.3 / -16.2
en→id	80.8 / 15.4 / +65.4	44.4 / 55.6 / -11.2	37.5 / 57.8 / -20.3	48.8 / 43.9 / +4.9
en→is	94.5 / 5.5 / +89.0	--	84.1 / 13.5 / +70.6	34.9 / 61.9 / -27.0
en→ja	83.5 / 13.4 / +70.1	39.3 / 55.7 / -16.4	31.5 / 52.2 / -20.7	29.9 / 55.9 / -26.0
en→kk	89.6 / 10.4 / +79.2	83.5 / 15.7 / +67.8	58.7 / 37.7 / +21.0	43.2 / 52.0 / -8.8
en→ko	86.6 / 11.0 / +75.6	44.5 / 48.4 / -3.9	31.0 / 61.9 / -30.9	43.2 / 48.0 / -4.8
en→lij	78.2 / 20.2 / +58.0	--	67.7 / 26.3 / +41.4	52.0 / 44.0 / +8.0
en→lld	70.2 / 27.4 / +42.8	--	79.7 / 17.6 / +62.1	53.7 / 39.7 / +14.0
en→ru	72.4 / 13.0 / +59.4	51.2 / 32.8 / +18.4	33.7 / 34.1 / -0.4	35.7 / 40.5 / -4.8
en→th	94.9 / 5.1 / +89.8	58.0 / 35.8 / +22.2	34.6 / 54.6 / -20.0	66.7 / 26.9 / +39.8
en→uk	87.2 / 6.4 / +80.8	46.8 / 46.0 / +0.8	38.6 / 52.6 / -14.0	29.5 / 57.4 / -27.9
en→zh	62.7 / 26.2 / +36.5	46.0 / 37.1 / +8.9	41.0 / 41.4 / -0.4	30.4 / 47.2 / -16.8
en→zh-Hant	61.0 / 30.9 / +30.1	48.0 / 35.4 / +12.6	37.5 / 47.8 / -10.3	33.1 / 48.8 / -15.7
zh→ja	85.9 / 9.4 / +76.5	50.8 / 35.9 / +14.9	42.5 / 37.2 / +5.3	45.7 / 40.2 / +5.5
Overall	83.2 / 13.7 / +69.5	52.4 / 40.8 / +11.6	52.3 / 38.3 / +14.0	38.7 / 52.4 / -13.7

IFMTBench multi-constraint (by direction)

Direction	Qwen3.5-9B W/L/G	HY-MT2-7B W/L/G	Gemma4-12B W/L/G	DeepSeek-v4-flash W/L/G
de→en	75.0 / 0.0 / +75.0	37.5 / 37.5 / 0.0	12.5 / 37.5 / -25.0	25.0 / 25.0 / 0.0
en→de	63.6 / 27.3 / +36.3	54.5 / 45.5 / +9.0	18.2 / 72.7 / -54.5	18.2 / 63.6 / -45.4
en→ja	64.3 / 21.4 / +42.9	40.0 / 33.3 / +6.7	53.3 / 26.7 / +26.6	28.6 / 35.7 / -7.1
en→ko	88.9 / 11.1 / +77.8	33.3 / 55.6 / -22.3	55.6 / 22.2 / +33.4	22.2 / 77.8 / -55.6
en→zh	43.5 / 34.8 / +8.7	42.6 / 31.3 / +11.3	30.8 / 29.1 / +1.7	23.1 / 41.0 / -17.9
ja→en	50.0 / 8.3 / +41.7	16.7 / 33.3 / -16.6	33.3 / 25.0 / +8.3	8.3 / 33.3 / -25.0
ja→zh	65.6 / 6.2 / +59.4	56.2 / 14.6 / +41.6	50.0 / 14.6 / +35.4	36.5 / 27.1 / +9.4
ko→en	64.3 / 14.3 / +50.0	21.4 / 35.7 / -14.3	21.4 / 35.7 / -14.3	42.9 / 21.4 / +21.5
ko→zh	47.4 / 36.8 / +10.6	42.1 / 26.3 / +15.8	36.8 / 31.6 / +5.2	26.3 / 36.8 / -10.5
zh→en	43.3 / 29.4 / +13.9	56.9 / 18.1 / +38.8	28.0 / 25.1 / +2.9	31.6 / 36.6 / -5.0
zh→ja	84.6 / 15.4 / +69.2	46.2 / 30.8 / +15.4	30.8 / 46.2 / -15.4	38.5 / 30.8 / +7.7
zh→ko	83.3 / 16.7 / +66.6	50.0 / 38.9 / +11.1	38.9 / 55.6 / -16.7	44.4 / 44.4 / 0.0
Overall	52.5 / 26.6 / +25.9	45.9 / 28.5 / +17.4	31.9 / 30.8 / +1.1	28.3 / 38.8 / -10.5

Prompts

Use the English name from the language table below for {src_lang} / {tgt_lang} (not the abbreviation). Separate paragraphs in {text} with blank lines (\n\n).

Translate the following {src_lang} text into {tgt_lang}. Strictly follow the instruction below. Output only the translation.

instruction:
{instruction}

text:
{text}

Use with transformers

Requires a recent transformers with Qwen3.5 support (transformers>=5.2.0 recommended).

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name_or_path = "WPS-Qingqiu/Qingqiu-MT-9B"

tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_name_or_path,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)

PROMPT = (
    "Translate the following {src_lang} text into {tgt_lang}. "
    "Strictly follow the instruction below. Output only the translation.\n\n"
    "instruction:\n{instruction}\n\n"
    "text:\n{text}"
)

messages = [
    {
        "role": "user",
        "content": PROMPT.format(
            src_lang="English",
            tgt_lang="Simplified Chinese",
            instruction="Use a formal register suitable for an official notice. Output only the translation.",
            text="Campus dining hall hours will change next week.\n\nBreakfast starts at 7:00.",
        ),
    }
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,  # non-thinking mode
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=2048,
    do_sample=True,
    temperature=0.3,
    top_p=0.90,
    top_k=20,
    repetition_penalty=1.05,
)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True).strip())

We recommend using the following set of parameters for inference:

{
  "top_k": 20,
  "top_p": 0.90,
  "repetition_penalty": 1.05,
  "temperature": 0.3
}

Supported languages

Languages	Abbr.	Chinese Names
English	en	英语
Simplified Chinese	zh	简体中文
Traditional Chinese	zh-Hant	繁体中文
Czech	cs	捷克语
Japanese	ja	日语
Korean	ko	韩语
German	de	德语
Russian	ru	俄语
Ukrainian	uk	乌克兰语
Vietnamese	vi	越南语
Indonesian	id	印度尼西亚语
Thai	th	泰语
Estonian	et	爱沙尼亚语
Icelandic	is	冰岛语
Belarusian	be	白俄罗斯语
Armenian	hy	亚美尼亚语
Kazakh	kk	哈萨克语
Egyptian Arabic	arz	埃及阿拉伯语
Ligurian (Italy)	lij	利古里亚语
Ladin (Italy)	lld	多洛米蒂拉定语

Main WMT26 directions:
cs→de/uk/vi, en→arz/be/cs/de/et/hy/id/is/ja/kk/ko/lij/lld/ru/th/uk/zh/zh-Hant, zh→ja.

Citation

@inproceedings{xia2026qingqiu,
  title={Qingqiu-MT-9B: An Instruction-Following Multilingual Translation Model},
  author={Xia, Tian and Chen, Chao and Yang, Mengpeng and Yang, Jingxu and Sun, Yabo and Liu, Qiang},
  booktitle={Proceedings of the Eleventh Conference on Machine Translation (WMT)},
  year={2026}
}

Authors & contact

Tian Xia, Chao Chen, Mengpeng Yang, Jingxu Yang, Yabo Sun, Qiang Liu
Kingsoft Office WPS-Qingqiu / Wuhan, China

xiatian3@wps.cn, chenchao9@wps.cn, yangmengpeng@wps.cn, yangjingxu1@wps.cn, sunyabo@wps.cn, liuqiang2@wps.cn

Built on Qwen3.5-9B by Alibaba Cloud.