SEA-LION is a collection of Large Language Models (LLMs) which have been pretrained and instruct-tuned for the Southeast Asia (SEA) region.
Qwen-SEA-LION-v4-4B-VL is a 4-billion parameter Vision-Language Model (VLM) built upon the Qwen3-VL-4B-Instruct architecture. To ensure domain adaptation for the region, the model underwent rigorous supervised fine-tuning (SFT) on a curated dataset of approximately 9 million instruction-text pairs. This extensive post-training instills multilingual and multicultural fluency, covering English and 7 key SEA languages: Burmese, Indonesian, Filipino, Malay, Tamil, Thai, and Vietnamese.
Qwen-SEA-LION-v4-4B-VL inherits the following features from Qwen3-VL:
SEA-LION stands for Southeast Asian Languages In One Network.
We performed post-training in English and SEA languages on Qwen3-VL-4B-Instruct, a decoder model using the Qwen3 architecture, to create Qwen-SEA-LION-v4-4B-VL.
For tokenization, the model employs the default tokenizer used in Qwen3-VL.
Developed by: AI Products Pillar, AI Singapore
Funded by: Singapore NRF
Shared by: AI Products Pillar, AI Singapore
Model type: Decoder
Context length: 256k
Language(s): fine-tuned on Burmese, Indonesian, Filipino, Malay, Tamil, Thai, and Vietnamese
The model has not been aligned for safety. Developers and users should perform their own safety fine-tuning and related security measures. In no event shall the authors be held liable for any claims, damages, or other liabilities arising from the use of the released weights and codes.
Bias, Risks, and Limitations
The model was not tested for robustness against adversarial prompting. It is important for users to be aware that our model exhibits certain limitations that warrant consideration. Like many LLMs, the model can hallucinate and occasionally generates irrelevant content, introducing fictional elements that are not grounded in the provided context. Users should also exercise caution in interpreting and validating the model's responses due to the potential inconsistencies.
How to Get Started with the Model
Use the code below to get started with the model with 🤗 Transformers libraries.
pip install transformers>=4.57.0
from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
# default: Load the model on the available device(s)
model = Qwen3VLForConditionalGeneration.from_pretrained(
"aisingapore/Qwen-SEA-LION-v4-4B-VL", dtype="auto", device_map="auto"
)
# We recommend enabling flash_attention_2 for better acceleration and memory saving, especially in multi-image and video scenarios.
# model = Qwen3VLForConditionalGeneration.from_pretrained(
# "aisingapore/Qwen-SEA-LION-v4-4B-VL",
# dtype=torch.bfloat16,
# attn_implementation="flash_attention_2",
# device_map="auto",
# )
processor = AutoProcessor.from_pretrained("aisingapore/Qwen-SEA-LION-v4-4B-VL")
messages = [
{
"role": "system",
"content": [{"type": "text", "text": "You are a helpful assistant."}]
},
{
"role": "user",
"content": [
{"type": "text", "text": "Write a poem on southeast asian countries in Indonesian."}
],
}
]
# Preparation for inference
inputs = processor.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
return_dict=True,
return_tensors="pt"
)
inputs = inputs.to(model.device)
# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=128)
generated_ids_trimmed = [
out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)
Training Details
Training Data
The instruction fine-tuning text dataset comprises of a collection of OSS & synthetic data.
Training Procedure
Training Hyperparameters
Training regime: Our workflow consists of instruction fine-tuning and model merging.
Evaluation
Testing Data, Factors & Metrics
We evaluated Qwen-SEA-LION-v4-4B-VL on general language, multi-turn chat and instruction-following capabilities.
Testing Data
General language capabilities
For the evaluation of general language capabilities, we employed the SEA-HELM evaluation benchmark across a variety of tasks. These tasks include Question Answering (QA), Sentiment Analysis (Sentiment), Toxicity Detection (Toxicity), Translation in both directions (Eng>Lang & Lang>Eng), Abstractive Summarisation (Abssum), Causal Reasoning (Causal), Natural Language Inference (NLI), Linguistic Diagnostics (LINDSEA), Cultural Knowledge (Kalahi) and Global MMLU Lite.
Instruction-following and Multi-turn Chat
We evaluated the models on instruction-following and multi-turn chat capabilities with SEA-IFEval (based on IFEval) and SEA-MTBench (based on MT-Bench) respectively. The two datasets were originally in English, the linguists and native speakers in the team worked together to filter, localise and translate the datasets into the respective target languages to ensure that the examples remained reasonable, meaningful and natural.
Factors
All evaluations were run with the model specific generation parameters defined in the model config. Each evaluation comprised of 8 runs with different seeds and the final results were averaged across these runs.
For all tasks, the model was expected to provide an answer tag from which the answer was automatically extracted. For tasks where options were provided, the answer should comprise one of the pre-defined options.
The evaluation was done zero-shot with native prompts on a sample of 100-1000 instances for each dataset.
SEA-IFEval
SEA-IFEval evaluates a model's ability to adhere to constraints provided in the prompt, for example beginning a response with a specific word/phrase or answering with a certain number of sections. Additionally, accuracy is normalised by the proportion of responses in the correct language (if the model performs the task correctly but responds in the wrong language, it is judged to have failed the task).
SEA-MTBench
SEA-MTBench evaluates a model's ability to engage in multi-turn (2 turns) conversations and respond in ways that align with human needs. We use gpt-4.1-2025-04-14 as the judge model and compare against gpt-4.1-2025-04-14 as the baseline model. The metric used is the weighted win rate against the baseline model (i.e. average win rate across each category: Math, Reasoning, STEM, Humanities, Roleplay, Writing, Extraction).
Metrics
The following metrics were used for text capabilities:
We also evaluated our models on two types of tasks using datasets specifically focused on Southeast Asian examples to benchmark and compared our models' performances against the original base models (Qwen3-VL-4B/8B).
Visual Question Answering (VQA): We utilised Multiple Choice Question (MCQ) style tasks, including MARVL, CVQA, and WorldCuisines.
Image Captioning: We employed the XM3600 dataset, evaluating strictly on examples relevant to the SEA region.
Key Insight: Despite our fine-tuning process focusing primarily on text data (approximately 9 million regional Q&A and instruction pairs), our evaluations confirm that Qwen-SEA-LION-v4 (4B/8B) successfully retains the high-performance vision-language capabilities of the original base models.
Factors
The evaluation was done zero-shot with native prompts.
Metrics
The following metrics were used to measure performance:
Normalised accuracy was the primary metric for the VQA tasks (CVQA, MARVL, and WorldCuisines).
RefCLIP Score was used for the XM3600 image captioning task.
Results
VL_result!
More Information
This is the repository for the commercial instruction-tuned model. The model has not been aligned for safety. Developers and users should perform their own safety fine-tuning and related security measures. In no event shall the authors be held liable for any claims, damages, or other liabilities arising from the use of the released weights and codes.
Ahn Jeongmi, Antonyrex Sajeban, Chan Hok Teng Adwin, Cheng Zi Yi Nicholas, Choa Hsueh Mei Esther, Heng Jonathan, Huang Yuli, Jann Railey Estrada Montalan, Lau Wayne, Lee Chwan Ren, Leong Wai Yi, Leong Wei Qi, Liew Rachel, Limkonchotiwat Peerat, Muhammad Ridzuan Bin Mokhtar, Nagarajan Karthik, Ng Boon Cheong Raymond, Ngee Chia Tai, Ngui Jian Gang, Nguyen Thanh Ngan, Ong Tat-Wee David, Ong Zhi Hao, Pereira Mark, Poon Joseph, Rengarajan Hamsawardhini, Siow Wei Kang Bryan, Susanto Yosephine, Sutaveephamochanon Anocha, Tan Choon Meng, Tan Chor Phin Evelyn, Tan Siao Wei Jessica, Tan Yixian, Tee Jun Yun, Teng Kok Wai Walter, Teo Eng Sipp Leslie, Tjhi William, Yeo Yeow Tong, Yong Xianbin, Zhang Zhou
Acknowledgement
This project is supported by the National Research Foundation Singapore and Infocomm Media Development Authority (IMDA),
Singapore under its National Large Language Model Funding Initiative.
