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

Libraries

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

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

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

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("96kevinli29/Qwen3-4B-SFT-Math-Multilingual")
model = AutoModelForCausalLM.from_pretrained("96kevinli29/Qwen3-4B-SFT-Math-Multilingual", device_map="auto")
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=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

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

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "96kevinli29/Qwen3-4B-SFT-Math-Multilingual"
# 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-SFT-Math-Multilingual",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

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

SGLang

How to use 96kevinli29/Qwen3-4B-SFT-Math-Multilingual 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-SFT-Math-Multilingual" \
    --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-SFT-Math-Multilingual",
		"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-SFT-Math-Multilingual" \
        --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-SFT-Math-Multilingual",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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

Qwen3-4B-SFT-Math-Multilingual

Qwen3-4B that reasons in the language of the problem

$Bar chart, non-English mean. Reasoning in the requested language: released 40%, SFT 99%, RL additive 100%, RL gated 100%. Did not finish: 5%, 21%, 1%, 2%. Delivered: 29%, 60%, 60%, 61%.$

_{Mean over Chinese, Spanish, French, Arabic and Russian · 70 problems × 16 samples per language}

What it is. Qwen3-4B after full-parameter fine-tuning on SFT-Math-Multilingual-43k: 43,218 long chain-of-thought math solutions in six languages. Epoch-2 checkpoint.
What it fixes. The released model solves Spanish, French and Arabic problems but writes its reasoning in English. This one reasons in the requested language: 39.8% → 99.4% of non-English responses, and best-of-16 delivery goes from 35.7% to 84.6%.
What to use it for. A warm start for RL. On its own, 21.1% of non-English responses fall into a repetition loop; Qwen3-4B-MathRL-Multilingual-Gated and Qwen3-4B-MathRL-Multilingual-Additive remove that.

Results

	Released	SFT	RL additive	RL gated
Answer correct	76.2%	61.4%	60.1%	61.7%
Reasoning in the requested language	39.8%	99.4%	99.6%	99.7%
Did not finish	4.7%	21.1%	1.2%	1.8%
Delivered (J@1)	28.6%	60.2%	59.8%	61.4%
Delivered, best of 16 (J@16)	35.7%	84.6%	83.1%	83.7%
Median response, tokens	5,824	7,673	4,072	4,463

70 competition problems (30 AIME-2026 + 40 AMC-2023) per language from Multilingual-Math-Reasoning-Eval, 16 samples each at temperature 0.7, 24,576-token budget. A response is delivered when the answer is correct, the visible reasoning is in the requested language, and it terminates (J = C ∧ L ∧ T, scored per response). Tables show the mean over Chinese, Spanish, French, Arabic and Russian.

Per language · other benchmarks

	Correct	In language	Did not finish	J@1	J@16	Median tokens
English	71.8%	100.0%	3.2%	71.7%	94.3%	5,369
Chinese	57.5%	99.9%	20.8%	57.3%	85.7%	5,946
Spanish	64.8%	99.6%	12.3%	64.6%	87.1%	7,044
French	64.9%	99.0%	12.9%	64.6%	87.1%	7,512
Arabic	56.1%	98.4%	31.7%	53.6%	80.0%	9,103
Russian	63.9%	99.9%	27.5%	61.2%	82.9%	8,760

Delivered (J), non-English mean	Released	SFT	RL additive	RL gated
PolyMath, 444 problems	16.8%	34.1%	35.4%	36.8%
MMLU-ProX-Lite, 588 questions	24.0%	49.0%	56.2%	55.5%

Quick start

from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer

name = "96kevinli29/Qwen3-4B-SFT-Math-Multilingual"
tok = AutoTokenizer.from_pretrained(name)
model = AutoModelForCausalLM.from_pretrained(name, torch_dtype="auto", device_map="auto")

# a French problem with a French instruction: "…en raisonnant entièrement en français."
prompt = load_dataset("96kevinli29/Multilingual-Math-Reasoning-Eval", "fr")["amc23"][0]["prompt"]
text = tok.apply_chat_template([{"role": "user", "content": prompt}], tokenize=False, add_generation_prompt=True)
inputs = tok(text, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=24576, do_sample=True, temperature=0.7)
print(tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

Qwen3 chat template with thinking on; the reasoning appears inside <think> … </think>. Give hard problems a large max_new_tokens. Loads with transformers 4.x and 5.x.

Prompt format in six languages

	Instruction that opens the user turn
English	Solve the following math problem step by step, reasoning entirely in English.
Chinese	请全程用中文逐步推理并解决以下数学问题。
Spanish	Resuelve el siguiente problema matemático paso a paso, razonando enteramente en español.
French	Résous le problème de mathématiques suivant étape par étape, en raisonnant entièrement en français.
Arabic	حل المسألة الرياضية التالية خطوة بخطوة مع التفكير بالكامل باللغة العربية.
Russian	Решите следующую математическую задачу шаг за шагом, рассуждая полностью на русском языке.

The full prompts (instruction, problem and answer format, all in the target language) are in the evaluation set.

Model family

Stage	Checkpoint
I · Release	Qwen/Qwen3-4B · Qwen/Qwen3-8B	solves the problem, reasons in English
II · SFT	Qwen3-4B-SFT-Math-Multilingual ← this model	reasons in the requested language, sometimes loops
II · SFT	Qwen3-8B-SFT-Math-Multilingual	the same result at 8B
III · RL	Qwen3-4B-MathRL-Multilingual-Gated	keeps the language, stops the looping
III · RL	Qwen3-4B-MathRL-Multilingual-Additive	keeps the language, stops the looping

Data: SFT-Math-Multilingual-43k · SFT-Math-Multilingual-56k · DAPO-Math-Multilingual-6Lang · Multilingual-Math-Reasoning-Eval

Training

Each of 45K English OpenR1-Math traces is assigned one of six languages, translated by a locally served Qwen3-14B and filtered by quality gates. 5 epochs configured; epoch 2 is released, where held-out loss has flattened and looping is still moderate. Code: base-model-sft-verl.

Full recipe


Objective	token-level cross-entropy over the whole response, reasoning included
Optimizer	AdamW, lr 2e-5 with cosine decay to 0.1×, warmup ratio 0.01, weight decay 0.1
Batch · context	16 traces · up to 32,768 tokens
Precision · clip	bfloat16 · gradient clip 1.0
Hardware	16 × A100-40GB on 4 nodes, MeluXina (LuxProvide)
Framework	verl SFT trainer, FSDP, full-parameter

Limitations

Looping: 21.1% of non-English responses do not finish within 24,576 tokens (31.7% in Arabic).
Accuracy is below the released Qwen3-4B on the same problems (76.2% → 61.4%). The same drop appears after English-only SFT, so it is not caused by the multilingual mixture.
Mathematics and six languages only; machine-translated training data; one seed.

Citation

@article{li2026posttraining,
  title   = {What Does Post-Training Change in Multilingual Reasoning?},
  author  = {Li, Hongyang and Li, Xiao and Wu, Caesar and Danoy, Gr{\'e}goire and Bouvry, Pascal},
  journal = {arXiv preprint arXiv:2609.37104},
  year    = {2026}
}

@misc{qwen3-4b-sft-math-multilingual-2026,
  title        = {{Qwen3-4B-SFT-Math-Multilingual}: six-language long chain-of-thought math SFT of {Qwen3-4B}},
  author       = {Li, Hongyang and Li, Xiao and Wu, Caesar and Danoy, Gr{\'e}goire and Bouvry, Pascal and {Sea-Fill Community}},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/96kevinli29/Qwen3-4B-SFT-Math-Multilingual}},
  note         = {Maintained by Sea-Fill Community.}
}

Built and released by the authors together with the Sea-Fill Community, which maintains this release.

Downloads last month: 51

Safetensors

Model size

4B params

Tensor type

BF16

Model tree for 96kevinli29/Qwen3-4B-SFT-Math-Multilingual

Base model

Qwen/Qwen3-4B-Base

Finetuned

Qwen/Qwen3-4B

Finetuned

(1,273)

this model

Finetunes

2 models

Datasets used to train 96kevinli29/Qwen3-4B-SFT-Math-Multilingual

Paper for 96kevinli29/Qwen3-4B-SFT-Math-Multilingual

What Does Post-Training Change in Multilingual Reasoning?

Paper • 2609.37104 • Published Sep 29

Evaluation results

Delivered J@1 (correct, in the requested language, terminated) on AIME-2026 + AMC-2023, five non-English languages (zh, es, fr, ar, ru)
self-reported

60.200
Answer accuracy on AIME-2026 + AMC-2023, five non-English languages (zh, es, fr, ar, ru)
self-reported

61.400
Reasoning in the requested language on AIME-2026 + AMC-2023, five non-English languages (zh, es, fr, ar, ru)
self-reported

99.400