SEA-LION
AI Singapore's open-weight LLMs tuned for Southeast Asian languages
Overview
SEA-LION is AI Singapore's family of open large language models for Southeast Asian languages and cultures. The latest release, v4.8 (September 2026), adds 30B and 120B mixture-of-experts models built with NVIDIA on Nemotron 3, MIT-licensed with a 262K-token context. Weights are free on Hugging Face, and a rate-limited free API serves proofs of concept.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 19 Oct 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
- v4.8 and v4.5 models MIT-licensed
- Free for commercial use, per the FAQ
- FP8, NVFP4 and GGUF variants of v4.8
- 10 requests per minute per user
- One key per user, Google sign-in
- OpenAI-compatible endpoints; limit cannot be raised
- Guides for Google Vertex AI, Amazon Bedrock and SageMaker
- Gemma-SEA-LION-v4-27B-IT on Cloudflare Workers AI
- Self-hosting with vLLM
Use cases
What you can produce with SEA-LION
- Free downloadable weights from the 2B-class Gemma E2B model up to the 120B-A12B Nemotron model
- An OpenAI-compatible trial API key for proofs of concept, capped at 10 requests a minute
- Chat and instruction following in Burmese, Filipino, Indonesian, Malay, Tamil, Thai and Vietnamese, plus English
- Image-and-text understanding from the v4 vision-language models
- Content moderation tuned to Southeast Asian norms with the SEA-Guard safety models
- Multilingual embeddings for search and retrieval from the SEA-LION embedding models
- Evaluations against the SEA-HELM benchmark suite and its public leaderboard
ASEAN Perspective
SEA-LION in Southeast Asia
SEA-LION is the only government-backed, purpose-built open LLM for the ASEAN region, developed in Singapore under a S$70 million National Multimodal LLM Programme and covering 11 official Southeast Asian languages. Its training data explicitly targets linguistic code-switching, regional colloquialisms, and cultural context that general-purpose Western LLMs systematically under-serve — a critical gap for Indonesian, Thai, Vietnamese, and Filipino enterprise and public-sector deployments. From a data-residency standpoint, self-hosted weights can be run entirely within-country (Indonesia, Thailand, Singapore, Malaysia) with no cross-border data transfer, making SEA-LION attractive under PDPA, Thailand PDPA, and Indonesia's PDP Law. Enterprise adoption by the GoTo Group ecosystem (Indonesia), NCS (Singapore), and IBM's ASEAN watsonx practice confirms real-world traction beyond government-funded pilots.
SEA-LION is the open model family to start with for Southeast Asian language work: regularly updated, MIT-licensed in its newest releases, and evaluated on SEA-HELM, the regional benchmark its own team maintains. The September 2026 v4.8 models bring mixture-of-experts efficiency and a 262K context, with large reported gains in Burmese and Tamil.
Read the limits before committing. The free API is capped at 10 requests a minute and meant for proofs of concept only, the newest model card lists seven regional languages rather than eleven, and the benchmark numbers come from the developer. Production means self-hosting or a cloud partner at your own cost.
What we found
SEA-LION (Southeast Asian Languages in One Network) is AI Singapore's family of open large language models for the region. The first models came out in December 2023; earlier versions of this page gave October. Since then AI Singapore has shipped v2, v3, v3.5, v4, v4.5 and, on 18 September 2026, v4.8: two Nemotron-SEA-LION models built with NVIDIA on the Nemotron 3 Nano and Super bases, at 30 billion and 120 billion parameters with 3 billion and 12 billion active. They are the series' first mixture-of-experts and first Nemotron-based models, use a Mamba2-Transformer hybrid with a 262,144-token context, and are MIT-licensed, with FP8, NVFP4 and GGUF builds for smaller hardware.
Language coverage needs reading closely. The 120B model card lists English plus seven regional languages: Burmese, Filipino, Indonesian, Malay, Tamil, Thai and Vietnamese. AI Singapore's FAQ names Khmer and Lao among the languages SEA-LION supports, but its SEA-HELM documentation says the benchmark does not yet cover them, so the homepage's "11+ languages" is broader than the evidence for any single model. On SEA-HELM, AI Singapore reports the 120B model lifting the Nemotron baseline from 49.30 to 63.44, with Burmese rising from 4.98 to 31.35. Those are the developer's figures on the developer's own benchmark, whose code reached version 1.3.0 in August 2026.
Trying it is free. A Google sign-in at the playground issues one trial API key, OpenAI-compatible and capped at 10 requests a minute; the FAQ says that tier is for proofs of concept and the limit cannot be raised. Production means self-hosting the weights or using a cloud: the docs cover Google Vertex AI, Amazon Bedrock, SageMaker, vLLM and Cloudflare Workers AI. Licences follow each model's Hugging Face tag, and some Gemma- and Llama-based releases carry those vendors' terms.
Development is funded through Singapore's S$70 million National Multimodal LLM Programme, a two-year National Research Foundation initiative launched in December 2023 that also covers A*STAR's MERaLiON. For ASEAN teams SEA-LION is a credible open starting point, not a turnkey service.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
Notable facts
- SEA-LION stands for Southeast Asian Languages In One Network — the acronym is a play on the Merlion, Singapore's iconic lion-fish national symbol.
- The v4.5 speculative decoder achieves up to 5× throughput by drafting tokens in parallel using a custom block-diffusion technique trained specifically on SEA text.
- SEA-LION's training corpus includes Javanese and Sundanese — two of Java's regional languages — making it one of very few LLMs to handle sub-national Indonesian linguistic diversity.
- The companion SEALD dataset project (co-led with Google Research and universities in Thailand and the Philippines) is building Southeast Asia's largest open multilingual text-and-speech corpus.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including SEA-LION's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with SEA-LION unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to SEA-LION directly →
Spotted something out of date? Suggest an update →
SEA-LION in the news
Alternatives to SEA-LION
More in LLMs & Chat
