Llama-SEA-LION-v3-70B-IT
AI Singapore's 70B Southeast Asian multilingual model: Llama 3.1 continued-pretrained on regional data, 128K context, 11 SEA languages.
About
Llama-SEA-LION-v3-70B-IT is the largest model in AI Singapore's SEA-LION v3 family, released December 2024. It takes Llama 3.1 70B and continues pretraining on Southeast Asian data, then instruction-tunes it with a mix of English and ASEAN-language instruction pairs, with the 9B sibling's recipe using roughly 500K English and a million regional-language pairs. The v3 generation covers 11 official Southeast Asian languages, including lower-resource ones like Javanese and Sundanese, and AI Singapore evaluates the family on SEA-HELM, its own regional benchmark suite.
The 70B instruction-tuned variant carries a 128K context window, unusual for regional models, and the pretraining used 200B high-quality SEA tokens on SingTel H100 clusters. AI Singapore has since moved on to v4 (Gemma 3-based, multimodal) and a Nemotron-based v4.8, so v3 is no longer the cutting edge of the line, but it remains the biggest openly available SEA-focused model. Licensing follows the Llama 3.1 Community License, which is free to use but not OSI-open, so read the terms if you plan commercial deployment.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- AI Singapore: best-performing SEA-LION v3 model on SEA-HELM for its size class (provider claim, sea-lion.ai)
- SEA-HELM leaderboard is public at leaderboard.sea-lion.ai; no independent replication of the SOTA claim verified
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 6/10
Scored 6/10 (metrics tier, metrics-v1). Evidence: AI Singapore's SEA-LION v3 announcement (sea-lion.ai, Dec 2024) documents the 200B-token continued pretraining on 11 SEA languages and the SEA-HELM eval; HF repo aisingapore/Llama-SEA-LION-v3-70B-IT created Dec 11, 2024 (HF API). Open weights. Strengths: serious national-lab effort, transparent training recipe, public regional benchmark, 128K context. Docked for: (1) benchmark SOTA claims are AI Singapore's own; (2) tiny direct adoption (105 downloads) and the line has moved on to v4/v4.8; (3) Llama 3.1 Community License is not OSI-open. License: llama3.1 (Llama 3.1 Community License).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 4.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 7.0/10
- Manual review · sanity read + red-flag check
- 6.0/10
- · AI Singapore SEA-LION v3 announcement Dec 2024; 200B SEA-token continued pretraining, 11 languages
- · HF repo aisingapore/Llama-SEA-LION-v3-70B-IT created 2024-12-11; 105 downloads, 5 likes (HF API Oct 2026)
- · 128K context; SEA-HELM leaderboard public
- · Line moved on to v4/v4.5; Llama 3.1 Community License not OSI-open