Two Paths to Open, Small and Efficient AI: Announcing Apertus and Gemma SEA-LION v4 Models

ByAISG | AI Products

We are expanding the SEA-LION v4 family to include small, efficient, fully open model as well as regionally-aware Vision-Language (VL) model.

Today, the SEA-LION v4 family expands with two of our newest small models:

Gemma-SEA-LION-v4-4B-VL (The Small but Mighty)

Built on Google’s Gemma 3, this 4B model further builds on the existing multilingual capabilities to serve the SEA region. Despite its small footprint, it incorporates our first attempts at adding tool calling and image-text extraction capabilities.

Apertus-SEA-LION-v4-8B-IT (The Open Standard)

Built on the fully open Swiss AI Apertus architecture and dataset, this collaboration improves on its 8B foundation for the SEA region. This is our effort to create a transparent and auditable model that incorporates an open end-to-end dataset and training pipeline.

Open Evaluation with SEA-HELM

In line with the theme of being open, we are making our evaluation suite SEA-HELM open with the latest update to encourage more rigorous and reproducible evaluation results for the SEA region.

Check out their performance on regional benchmarks below:

Bar chart showing the overall SEA average scores for various models, including SEA-LION v4, Qwen, and Gemma, with error margins displayed.

Tool calling

Beyond benchmarking our Gemma-SEA-LION-v4-4B-VL and Apertus-SEA-LION-v4-8B-IT models for SEA multilingual capabilities, we are also exploring benchmarking our models for specific capabilities. Here, using the Berkeley Function Calling Leaderboard (BFCL) V4 evaluation, we compared our models against Gemma-3-4B-IT and Apertus-8B-Instruct-2509, both of which did not natively support tool calling:

Visual parsing

Given the strong multilingual performance of our Gemma-SEA-LION-v4-4B-VL, we also wanted to evaluate its visual parsing capabilities in English, Mandarin and Thai using the components of the OCRBenchv2 and ThaiOCRBench that focused primarily on text extraction, which includes the following tasks:

  • Table/Chart/Document parsing
  • Fine grained image-text extraction
  • Full page image-text extraction
  • Text recognition

We compared our Gemma-SEA-LION-v4-4B-VL against Gemma-3-4B-IT:

A table displaying the performance metrics of two models: Gemma-SEA-LION-v4-4B-VL and Gemma-3-4B-IT, including visual parsing scores for different languages (TH, ZH, EN).

Get Started Now

Both models are available for download today on HuggingFace and can be tested later in our playground.

We hope that these SEA-LION v4 models will suit your use case – whether for deployment to edge devices or as a fully open model solution.

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.