Qwen3-8B-ZH is a native reasoning model fine-tuned from Qwen/Qwen3-8B-Base to reason in Chinese. This model produces its entire reasoning trace in Chinese before delivering the final answer in Chinese.

It is released alongside the paper Rethinking the Multilingual Reasoning Gap with Layer Swap.

Model details

Base model: Qwen/Qwen3-8B-Base
Language: Chinese (CoT and answer)
Training: Full SFT, ~10B tokens, 2 epochs
Context length: 32,768 tokens
Dataset: lightonai/Dolci-Think-SFT-32B-Multilingual (Chinese split).

[!NOTE] The model was trained on data derived from allenai/Dolci-Think-SFT-32B, released under the ODC-BY-1.0 license.

Related models

This model is part of a Chinese specialist trio designed to study the native reasoning gap:

Model	CoT language	Description
`lightonai/Qwen3-8B-ZH`	Chinese	Native reasoning specialist
`lightonai/Qwen3-8B-ZH-Swap`	Chinese	Layer Swap: middle layers (L13–L20) of `Qwen3-8B-EN` transplanted into `Qwen3-8B-ZH`
`lightonai/Qwen3-8B-ZH-Pivot-EN`	English	Same Chinese Q&A pairs, but CoT in English
`lightonai/Qwen3-8B-EN`	English	English specialist

Evaluation

All scores are mean accuracy (%) on the Chinese version of each benchmark, with sample standard deviation across runs. AIME 24/25 is averaged over 30 runs; the others over 10 runs, using the recommended generation parameters.

Model	MGSM-Rev2	Global-MMLU-Lite	GPQA-Diamond	AIME 24/25	HumanEvalPlus	Average
`Qwen3-8B-ZH`	88.92	74.85	50.71	53.89	85.62	70.80
`Qwen3-8B-ZH-Swap`	88.24	76.42	52.58	55.17	85.69	71.62
`Qwen3-8B-ZH-Pivot-EN`	94.84	76.15	54.19	59.06	85.19	73.89
`Qwen3-8B-EN`	76.04	75.00	47.53	50.00	83.88	66.49

Benchmarks used:

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "lightonai/Qwen3-8B-ZH"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", device_map="auto")

messages = [{"role": "user", "content": "计算：24 × 17 = ?"}]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(model.device)

outputs = model.generate(inputs, max_new_tokens=32768, temperature=1.0, top_p=0.95, top_k=20)
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))

Recommended sampling: temperature=1.0, top_p=0.95, top_k=20, min_p=0.

Citation

If you find our work helpful, feel free to give us a cite.

bibtex

@misc{lasbordes2026rethinking,
  title        = {Rethinking the Multilingual Reasoning Gap with Layer Swap},
  author       = {Lasbordes, Maxence and Chatelain, Amélie and Seddah, Djamé},
  year         = {2026},
  eprint       = {2605.26735},
  archivePrefix= {arXiv},
  primaryClass = {cs.CL}
}