AI SingaporeSEA-LION v4

Gemma-SEA-LION-v4-27B-VL

Gemma-SEA-LION-v4-27B-VL adds regional image-text training to Gemma-SEA-LION-v4-27B-IT. AI Singapore and SEACrowd developed this vision-language release. Its training includes about 540,000 image-text instruction samples and a merge with Gemma 3 27B IT.

How this release works

The card lists regional visual-question-answering and captioning evaluations. It expects text capability to remain similar to the parent because this stage focuses on vision and text together. The release uses the Gemma 3 tokenizer and includes image-grounded reasoning and document comprehension.Gemma-SEA-LION-v4-27B-VL official model card

How to assess it for your work

Practical guidance. Assess this release on photographs, menus, forms and signs that contain local-language content. Keep the image quality fixed when comparing it with its parent. Inspect the evidence in each answer: a correct description of an image does not establish that all names, prices or dates were read correctly.

Specifications & access

Gemma-SEA-LION-v4-27B-VL specifications
DetailValue and conditions
AccessDownloadable
Model identifiersaisingapore/Gemma-SEA-LION-v4-27B-VL
Context stated by the card128K tokensGemma-SEA-LION-v4-27B-VL official model card
Input and outputImage and text inputs; text outputGemma-SEA-LION-v4-27B-VL official model card
License declared by the model cardMITGemma-SEA-LION-v4-27B-VL official model card
Parent checkpointGemma-SEA-LION-v4-27B-ITGemma-SEA-LION-v4-27B-VL official model card

Related models

  • A 27B Gemma 3 instruction model with Southeast Asian text training and retained image input support.

    Launch date not confirmedDownloadable
  • A 4B Gemma 3 vision-language model adapted for regional image understanding, visual parsing and tool calls.

    Launch date not confirmedDownloadable
  • An 8B Qwen3-VL model adapted with about nine million regional instruction-text pairs.

    Launch date not confirmedDownloadable