Sailor
Open multilingual language models for South-East Asia
Sea AI Lab created Sailor and Sailor2 to serve South-East Asian users with open language models. Built on Qwen bases, these systems support fifteen regional languages and range from small research sizes to larger production variants.
- Supports 15 languages including Thai, Vietnamese, and Javanese
- Released under Apache 2.0 license for open use
- Based on Qwen architecture with up to 20B parameters
- Trained on up to 500 billion tokens of regional data
- Developed by Sea AI Lab for South-East Asia
What it is
Sailor is a family of open language models designed specifically for South-East Asian contexts. Sea AI Lab developed the series to address the lack of robust tools for regional languages. The project offers models in various sizes, including 1 billion, 8 billion, and 20 billion parameters. These systems handle fifteen languages, covering major tongues like English and Chinese alongside local dialects such as Javanese and Tagalog. The goal is to provide accessible technology for both research and commercial use in the region.
How it came about
The initial Sailor models appeared in April 2024, ranging from 0.5 billion to 7 billion parameters. They were built by continually pre-training Qwen1.5 on hundreds of billions of tokens. In 2025, the team released Sailor2, expanding the family to include 20 billion parameter variants. This second generation extended the foundation to Qwen 2.5 and increased the training data volume. The evolution reflects a shift toward supporting larger production workloads while maintaining open access.
Inside the model
The models use a decoder-only transformer architecture derived from the Qwen series. Training involves processing between 200 billion and 500 billion tokens, depending on the version. Developers employed techniques like BPE dropout and aggressive data cleaning to improve robustness. The process includes multi-stage pre-training and optimized data mixing strategies. Post-training steps focus on multilingual alignment to ensure performance across all supported languages. The system uses a unified approach to handle diverse linguistic structures within a single model.
What it can do
Sailor models perform well on tasks involving commonsense reasoning and reading comprehension. They support question answering and examination-style queries in multiple languages. The 20 billion parameter variant aims for production-level accuracy, while smaller versions suit specialized research needs. Benchmarks indicate strong capabilities in handling Southeast Asian scripts and grammar. The models are designed to be faithful and informative when responding to user prompts in any of the fifteen supported tongues.
How to use it
All weights are available under the Apache 2.0 license, allowing free commercial and research use. Users can download the models from Hugging Face or access the codebase via GitHub. The project requires specific transformer library versions for optimal performance. There are no restrictions on deployment, making it suitable for startups and academic groups. The open nature encourages community contributions and further development of regional language tools.
Why it matters
Sailor has highlighted the demand for localized language technologies in South-East Asia. By providing open weights, it enables other developers to build applications for underserved communities. The project demonstrates that continual pre-training on regional data significantly improves performance over generic multilingual models. It serves as a reference for creating inclusive AI systems that respect linguistic diversity.
Questions people ask about Sailor
Which languages does Sailor support?
The model handles fifteen languages, including English, Chinese, Indonesian, Thai, Vietnamese, and Tagalog. It also supports less widely represented tongues such as Burmese, Lao, Javanese, and Waray, aiming for broad regional coverage.
Can I use Sailor for commercial products?
Yes, the models are distributed under the Apache 2.0 license. This permits both research and commercial usage without restrictive clauses, allowing developers to integrate the technology into paid services or internal tools freely.
What is the largest size available?
The Sailor2 family includes a model with 20 billion parameters. Smaller variants with 1 billion and 8 billion parameters are also available, catering to different computational constraints and application requirements.
Good reading and sources
There is no English Wikipedia article on Sailor yet, so the facts on this page come from the sources listed above: huggingface.co/sail/Sailor2-20B-Chat, arxiv.org/abs/2404.03608. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.