AI Singapore

AI Singapore (AISG) is Singapore's national AI programme, announced on 3 May 2017 with a commitment from the National Research Foundation to invest up to S$150 million over five years, co-driven by the Smart Nation and Digital Government Office, EDB, IMDA, SGInnovate, and IHiS, and hosted at the National University of Singapore. Its founding executive chairman was NUS deputy president Ho Teck Hua; the programme is now chaired by Christian Wolfrum and organised into pillars for AI Research, AI Technology, AI Innovation, AI Products, AI Governance, and Talent, with the 100 Experiments industry programme and the AI Apprenticeship Programme as its adoption arms. Its research identity is the AI Products pillar's SEA-LION project (Southeast Asian Languages In One Network), led by Senior Director Leslie Teo with Deputy Director William Chandra Tjhi.

SEA-LION began as a from-scratch effort: v1 (October 2023) pretrained 3B and 7B MPT-architecture models on 980B tokens with a custom 256K-entry Southeast Asian tokenizer on 256 A100s. From v2 (July 2024) the programme switched to continued pretraining and post-training of open foundation models, a choice the lab argues gets more capability per dollar for low-resource languages: v2 on Llama 3, v3 on Llama 3.1 and Gemma 2 with 200B tokens of CPT (December 2024, with a 70B), v4 on Gemma 3 27B with 500B tokens and the family's first vision input (August 2025, then Qwen3 32B and 8B-VL siblings), v4.5 on Qwen3.6-27B and Gemma 4 E2B with a block-diffusion speculative decoder (May 2026), and Nemotron-SEA-LION-v4.8 (September 2026), the family's first mixture-of-experts models, built with NVIDIA on Nemotron 3 Nano and Super with continued pretraining and on-policy distillation from Nemotron 3 Ultra. Every release ships under MIT (or the base model's terms) on HuggingFace, where the org holds about 90 models.

The lab's second contribution is evaluation for the region: SEA-HELM (2025, ACL Findings), built with Stanford CRFM's HELM team, scores models across NLP classics, LLM-specific tasks, Southeast Asian linguistics, culture, and safety in Filipino, Indonesian, Tamil, Thai, and Vietnamese, since extended to Malay and Burmese, and runs a public leaderboard the SEA-LION models are measured on. Companion work includes the SEA-LION Embedding suite (EMNLP 2026 Findings), SEA-SafeguardBench (ACL 2026 Findings), SEA-Guard safety models, and Project Aquarium's mapping of regional language data. Coverage spans 11 official Southeast Asian languages plus regional ones such as Javanese, Sundanese, and Tetum; no SEA-LION model is scored on Artificial Analysis.