Announcing Qwen-SEA-LION-v4 (4B & 8B): Additional Vision-Language Models for Southeast Asia

ByAISG | AI Products

Vision-Language Models (VLMs) are changing how AI perceives and describes the world—but does your AI see and explain the real Southeast Asia?

Adding to our family of models from Llama, Gemma and Qwen, we are proud to announce the release of small but powerful multilingual and multicultural text models. Like Gemma, these models also support visual inputs as an additional capability. Introducing: Qwen-SEA-LION-v4-4B-VL and Qwen-SEA-LION-v4-8B-VL.

Built on top of the Qwen3-VL architecture and fine-tuned specifically for our region, these models represent a significant leap forward in how small but performant AI models read, reason, and interact with Southeast Asian content.

Why Two Models?

We understand that one size doesn’t fit all. We also believe that high-performance AI shouldn’t require a supercomputer. To ensure our models are accessible even on low infrastructure, we are releasing two sizes tailored to your hardware constraints:

  • Qwen-SEA-LION-v4-4B-VL: A lightweight powerhouse engineered for maximum resource efficiency. It is optimised to run on consumer-grade hardware, including laptops, making it perfect for edge deployment and applications where speed and low latency are critical.
  • Qwen-SEA-LION-v4-8B-VL: Offers a step up in reasoning depth for complex tasks while maintaining a manageable footprint, allowing for robust performance without the need for massive industrial compute clusters.

SEA-LION’s Approach to Regional Visual Language Processing

Many global models face challenges when processing Thai script or visual contexts specific to regions like Jakarta. SEA-LION addresses these limitations.

These models have been instruction-tuned on a massive dataset of question-answer pairs specifically curated for Southeast Asian tasks. We’ve drilled down into 7 key regional languages:

  • Burmese
  • Indonesian
  • Filipino
  • Malay
  • Tamil
  • Thai
  • Vietnamese

Key Features that Matter

1. Proven Performance for SEA

We don’t just claim performance; we measure it. The models have been rigorously tested on the SEA-HELM evaluation benchmark across tasks ranging from tasks such as Sentiment Analysis, Translation, Toxicity Detection, to Linguistic Diagnostics and Cultural Diagnostics.

A bar chart displaying model performance averages for various AI models, including SEA-LION v4 and Qwen 3, with numerical values and error margins shown for comparison.

2. Massive Context for Massive Tasks

Forget checking the token count constantly. Both models inherit a native 256K context window.

  • Articles: Feed in entire PDF manuscripts.
  • Video: Analyze transcripts from hours-long videos.
  • Complex Docs: Reason across multi-page financial reports without losing the thread.

3. Competitive Image Captioning Capabilities

Image captioning, one of the vision capabilities inherited from Qwen3-VL, directly reflects how models perceive and describe the world. Whether it’s reasoning about food restrictions, describing a cultural site or explaining cultural practices, our models can capture the details other models miss.

Model in Action

Small but mighty. Real-world data is rarely clean. Watch the 8B-VL and 4B-VL models ignore distractions like heavy watermarks and emojis to capture exactly what matters in the image.

  1. Qwen-SEA-LION-v4-8B-VL: This image shows a food stall selling “Daun Pandan Rice Dumplings” (蘭葉香肉粽), a type of zongzi wrapped in pandan leaves and filled with meat. It is located in an open-air market or hawker center. The menu displays various dumpling fillings, such as pork, chicken, or mixed meats, all made from rice and meat — none contain dairy or milk products. Therefore, people with lactose intolerance can safely eat all of them.
A food stall showcasing 'Daun Pandan Rice Dumplings' displayed in colorful wrappers, with a sign highlighting the menu options. The setting appears to be an open-air market or food fair.
  1. Qwen-SEA-LION-v4-4B-VL: This image shows the entrance gate of a temple or religious site with traditional Vietnamese architecture. The building is two-storied with a curved roof, adorned with carvings and intricate decorations. The area around the building is well-maintained with trees and plants, and there are people walking toward the site, who may be tourists or visitors. The image conveys a peaceful and culturally significant atmosphere. The appropriate situation for this image is when tourists or visitors are entering a historically and culturally significant religious site for sightseeing or pilgrimage. The country where you can find this event is Vietnam.
This image shows the entrance gate of a temple in Vietnam, featuring traditional architecture with a curved roof and intricate carvings. Visitors are seen approaching the gate, surrounded by well-maintained greenery.
  1. Qwen-SEA-LION-v4-4B-VL: This image shows a group of women in traditional Thai attire, adorned with gold jewelry and other ornaments. They are wearing golden headdresses decorated with red flowers. Four women are standing in the back, and one is sitting in the front wearing a special Thai costume with a traditional Thai mask, which is characteristic of Thai classical dance or traditional performances. The photo was taken in a public place, possibly in front of a shop or shopping mall, with a sign reading “#01-11,” which is a store code. This situation could be taking place in Thailand, particularly in major cities like Bangkok or areas where cultural events or traditional performances are held, such as Thai classical dance or traditional Thai theater, which are part of Thailand’s rich cultural heritage promoted in daily life.
This image features a group of women in traditional Thai attire, adorned with golden jewelry and headdresses, standing alongside a performer in a colorful costume and mask, depicting a traditional Thai dance performance.

Try Them Today

The future of Southeast Asian AI is multimodal, and it’s available now.

Acknowledgements: This work is funded by the Singapore National Research Foundation (NRF) and developed by the AI Products Pillar at AI Singapore. For inquiries, please contact sealion@aisingapore.org.