SEA-LION-E5-Embedding-600M
SEA-LION-E5-Embedding-600M maps text into 1024-dimensional vectors for search and similarity work. It adapts multilingual-e5-large, which uses the XLM-RoBERTa architecture. The released weights focus on eleven languages used across the region.
How this release works
The listed languages are Burmese, Chinese, English, Filipino, Indonesian, Khmer, Lao, Malay, Tamil, Thai and Vietnamese. The card demonstrates Sentence Transformers with its STS prompt. It distinguishes the tuned embedding models from base encoder checkpoints intended for further training.SEA-LION-E5-Embedding-600M official model card
How to assess it for your work
Practical guidance. Use this model for search over short passages. Divide longer documents at useful content boundaries so a heading stays with the text it describes. Review retrieved passages before judging the answer produced by a separate language model. A retrieval failure and an answer-generation failure require different corrections.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | aisingapore/SEA-LION-E5-Embedding-600M |
| Context window | 512 tokens Maximum input context stated by the embedding model card. SEA-LION-E5-Embedding-600M official model card |
| Model size class | 600MSEA-LION-E5-Embedding-600M official model card |
| Vector dimensions | 1024SEA-LION-E5-Embedding-600M official model card |
| License declared by the model card | MITSEA-LION-E5-Embedding-600M official model card |
| Output | Text embeddingsSEA-LION-E5-Embedding-600M official model card |
Related models
- SEA-LION-ModernBERT-Embedding-300M
AI Singapore · SEA-LION Embedding
A 300M ModernBERT text embedding model with a custom tokenizer for Southeast Asian languages.
Launch date not confirmedDownloadable - SEA-LION-ModernBERT-Embedding-600M
AI Singapore · SEA-LION Embedding
A 600M ModernBERT text encoder for regional retrieval and similarity tasks with a custom Gemma 3 tokenizer.
Launch date not confirmedAPI · Downloadable