Instructions to use 96kevinli29/Qwen3-4B-Multilingual-Reasoning with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use 96kevinli29/Qwen3-4B-Multilingual-Reasoning with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="96kevinli29/Qwen3-4B-Multilingual-Reasoning")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("96kevinli29/Qwen3-4B-Multilingual-Reasoning")
model = AutoModelForCausalLM.from_pretrained("96kevinli29/Qwen3-4B-Multilingual-Reasoning")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use 96kevinli29/Qwen3-4B-Multilingual-Reasoning with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "96kevinli29/Qwen3-4B-Multilingual-Reasoning"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "96kevinli29/Qwen3-4B-Multilingual-Reasoning",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/96kevinli29/Qwen3-4B-Multilingual-Reasoning

SGLang

How to use 96kevinli29/Qwen3-4B-Multilingual-Reasoning with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "96kevinli29/Qwen3-4B-Multilingual-Reasoning" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "96kevinli29/Qwen3-4B-Multilingual-Reasoning",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "96kevinli29/Qwen3-4B-Multilingual-Reasoning" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "96kevinli29/Qwen3-4B-Multilingual-Reasoning",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use 96kevinli29/Qwen3-4B-Multilingual-Reasoning with Docker Model Runner:
```
docker model run hf.co/96kevinli29/Qwen3-4B-Multilingual-Reasoning
```

Qwen3-4B · Multilingual Reasoning — thinks in 6 languages

A Qwen3-4B whose chain-of-thought is written in the target language — English, Chinese, Spanish, French, Arabic, Russian — not just the final answer. Standard reasoning models silently think in English and translate only the result; this model produces its <think>...</think> reasoning natively in each language, making multilingual math reasoning transparent and language-consistent.

Model details

Base model: Qwen/Qwen3-4B (official post-trained), not Qwen3-4B-Base.
Fine-tuning: SFT on a balanced 6-language long chain-of-thought math set (SFT-Math-45k-mling6, 43k problems, one language per problem, round-robin), 2 epochs. (Internal name: Qwen3-4B-mling6-SFT-ep2.)
Architecture: Qwen3ForCausalLM, 4B params, 36 layers, hidden 2560, vocab 151,936, context 40,960.
Languages: English, Chinese, Spanish, French, Arabic, Russian.
Format: thinking model — emits a <think>...</think> trace, then the answer.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
name = "96kevinli29/Qwen3-4B-Multilingual-Reasoning"
tok = AutoTokenizer.from_pretrained(name)
model = AutoModelForCausalLM.from_pretrained(name, torch_dtype="bfloat16", device_map="auto")
msgs = [{"role": "user", "content": "Решите задачу шаг за шагом, рассуждая полностью на русском языке.\n\n..."}]
inputs = tok.apply_chat_template(msgs, add_generation_prompt=True, enable_thinking=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=24576, temperature=0.7)
print(tok.decode(out[0][inputs.shape[1]:]))

To get in-language reasoning, prompt the model in the target language (input language drives the reasoning language).

Notes & limitations

On some scripts (notably Arabic and Russian) genuine in-language reasoning can exhibit generation-stability issues (looping / not terminating within the token budget). Read accuracy together with generation-completion rate.
Companion RL prompt set for improving stability: 96kevinli29/DAPO-Math-Multilingual-6Lang.

Training data & attribution

Base model: Qwen/Qwen3-4B (official post-trained).
SFT data source: built by translating English long chain-of-thought traces from OpenR1-Math (open-r1/OpenR1-Math-220k) (via a 45k English subset) into the five non-English languages with Qwen3-14B (faithful, paragraph-chunked; LaTeX / \boxed{} / <think> tags preserved). English traces pass through unchanged. Please cite Qwen3 and OpenR1-Math as upstream sources.

Downloads last month: 189

Safetensors

Model size

4B params

Tensor type

BF16

Model tree for 96kevinli29/Qwen3-4B-Multilingual-Reasoning

Base model

Qwen/Qwen3-4B-Base

Finetuned

Qwen/Qwen3-4B

Finetuned

(869)

this model