From the archive
SEA-LION v3.5 adds reasoning and a choice of response modes
Regional language models gain switchable reasoning, alongside updated instruction-tuned releases.
AI Singapore released SEA-LION v3.5, adding reasoning models trained on Southeast Asian data. The launch includes eight-billion and 70-billion-parameter variants based on Meta's Llama 3.1.
Turn detailed reasoning on or off
The release supports a reasoning mode for more involved tasks and a concise mode for straightforward responses. The choice is controlled through the model's chat template. Both reasoning variants support a 128,000-token context window.
This gives applications a way to choose the response style around the task rather than always requesting the longest reasoning output.
Regional coverage continues to grow
The announcement describes support for thirteen languages, including Javanese and Sundanese. It also updates the earlier instruction-tuned v3 models with additional data and post-training work. The SEA-LION profile covers the model families and their downloadable or hosted access options.