Skip to main content
Rase AI Intelligence Newsletter

Models & assistants Model or AI service

SEA-LION v3 adapts Gemma and Llama for Southeast Asian languages

AI Singapore developed SEA-LION v3 language models for Southeast Asia. The published research describes Gemma- and Llama-based versions, including a 9B Gemma checkpoint trained further on regional languages. [2][3]

Model access
Downloadable checkpoints include an instruction-tuned Gemma 2 model and a separate Llama-based model. [2][3]
License
The Gemma checkpoint’s card lists MIT. Google’s Gemma terms also govern derivatives of the Gemma 2 base. [1][3]

What the model does

The team documented the Gemma 9B and Llama 8B instruction-tuned variants in an April 2025 preprint. Its Gemma model card describes further training and instruction tuning for regional use. [2][3]

Why it matters

Adapting an existing model can give teams a starting point for tasks in languages underrepresented in many general-purpose models. Researchers can compare the adapted model with its original base on the same examples. [2][3]

Who can use it

Developers working with languages such as Indonesian, Malay, Thai and Vietnamese can test the documented Gemma checkpoint. The research also covers the Llama-based version. [2][3]

Key dates

  1. SEA-LION v3 research first submitted [2]

Ways to use it

Evaluate a local-language document task using examples written by speakers of the language. Track which SEA-LION checkpoint you use and compare it with the corresponding unadapted base. [2][3]

Model limitations

AI Singapore lists different variants in the family. The Gemma checkpoint’s model card lists MIT and identifies its Gemma 2 base. Google publishes separate terms for use and distribution of Gemma derivatives. [1][3]

Sources and verification

Catalog evidence checked How we verify records

  1. Primary source Google’s Gemma Terms of Use covering Gemma 2 and derivatives
  2. Primary source AI Singapore’s first SEA-LION v3 technical report and submission history
  3. Primary source aisingapore/Gemma-SEA-LION-v3-9B-IT
  4. Other source SEA-LION: Southeast Asian Languages in One Network
  5. Primary source Hugging Face registry metadata for aisingapore/Gemma-SEA-LION-v3-9B-IT

Related profiles

Models & assistants

GPT-6.2 Astra and Sol

OpenAI

Astra targets demanding projects. Sol targets everyday tasks. Both support text and image input in the API.

High confidenceEvidence checked

Models & assistants

MAI Code 2 and Code 2 Flash

Microsoft AI

Code 2 targets complex software work. Flash targets frequent edits, explanations and routine fixes.

High confidenceEvidence checked

Models & assistants

MAI Realtime 1

Microsoft AI

Users can try a MAI Labs demo. Developers can register for a limited Foundry preview.

High confidenceEvidence checked