SEA-LION-v4.8: A Technical Report
Ahmed Mohammad Dabeer (David Wang Dawei), Ahn Jeongmi (David Wang Dawei), Anocha Sutaveephamochanon (David Wang Dawei), Antonyrex Sajeban (David Wang Dawei), Aulia Adila (David Wang Dawei), Chan Hok Teng (David Wang Dawei), Adwin (David Wang Dawei), Cheng Zi Yi (David Wang Dawei), et al.
cs.CL
Quick verdict
Adapts Nemotron 3 to SEA languages via CPT + online distillation with strong SEA-HELM gains (120B 49.30→63.44); solid engineering but evaluated only against its own base models.
Summary
This technical report from AI Singapore presents Nemotron-SEA-LION-v4.8, four open checkpoints (30B-A3B and 120B-A12B, each in base and post-trained variants) adapted from NVIDIA Nemotron 3. Continued pre-training uses 150B (30B) or 33.5B (120B) tokens of Southeast Asian instruction, reasoning, code, and bilingual parallel data; post-training interleaves supervised fine-tuning with online on-policy distillation (reverse KL against a 550B Nemotron-3-Ultra teacher) in a decoupled orchestration system sustaining ~3,000 concurrent agents and ≥100k interaction sessions, followed by merging with the original base weights. On the updated SEA-HELM benchmark (seven SEA languages, eight-run averaging with bootstrapping), overall scores improve from 46.06 to 51.57 (30B) and 49.30 to 63.44 (120B), with the largest gains in Tamil and Burmese. The authors candidly document limitations: the retained Nemotron tokenizer is roughly 5× less efficient for Khmer, Lao, and Tamil; Filipino has no dedicated CPT data; and cultural, knowledge, and safety capabilities show regressions in places. Key concerns are evaluation solely against the models' own bases on a benchmark curated by the same organization, unreported confidence intervals, no general-capability evaluation, and under-specified weight merging.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Continued pre-training on 150B (30B) / 33.5B (120B) SEA-centric tokens improves the overall SEA-HELM score from 46.06 to 51.57 (30B-A3B) and 49.30 to 63.44 (120B-A12B) over the corresponding Nemotron 3 bases.
- Post-training that interleaves SFT with online on-policy distillation (reverse KL from a 550B teacher on student-generated trajectories) produces the largest gains in instruction following, natural language reasoning, and natural language understanding across seven SEA languages.
- The adaptation improves every one of the seven evaluated SEA languages, with the largest gains in Tamil (120B: 21.83→56.35) and Burmese (120B: 4.98→31.35).
- Targeted regional adaptation via continued pre-training plus online post-training can substantially raise SEA language capability without training a new foundation model from scratch.
- Retaining the original Nemotron tokenizer imposes a fundamental bottleneck: Khmer, Lao, and Tamil require roughly five times more tokens than with the previous SEA-LION tokenizer, capping returns from additional training data.
Strengths
Unusually candid reporting of negative results
Sections 3.2-3.4 quantify the tokenizer bottleneck (~5× more tokens for Khmer/Lao/Tamil than the prior SEA-LION tokenizer) and explicitly admit Filipino is absent from the CPT corpus; Section 3.4 distinguishes the failure modes (representation-constrained Tamil/Burmese vs data-coverage-constrained Filipino).
Detailed, reproducibility-oriented infrastructure documentation
Table 4 and Section 4.2 specify reverse-KL OPD against the 550B-A55B Ultra teacher, staleness window of 32 tokens-steps, token-exact capture via Conductor, ~3,000 concurrent agents, ≥100k sessions, per-scale GPU counts (16/48 training, 48 reward, 8/16 serving on H200s) and runtimes (12h/24h).
Consistent language-level gains including tail languages
More rigorous evaluation protocol than typical model reports
Full-pipeline checkpoint release
Weaknesses
No third-party baselines
Statistical methodology described but not reported
General-capability preservation asserted but never measured
Weight merging is a material, under-specified confound
Pervasive internal inconsistencies suggest hasty preparation
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is SEA-LION-v4.8?
Nemotron-SEA-LION-v4.8 is a family of open large language models from AI Singapore focused on Southeast Asian languages, built on top of NVIDIA's Nemotron 3 foundation models. It comes in two sizes — 30B total parameters with 3B active (MoE) and 120B total with 12B active — and each size is released both as a continued-pretrained base model and as an instruction-following post-trained model, for four checkpoints in total.
How much better is SEA-LION-v4.8 than its base model?
On the SEA-HELM benchmark covering seven Southeast Asian languages, the 30B model's overall score improves from 46.06 to 51.57 and the 120B model's from 49.30 to 63.44. The largest gains are in Tamil (21.83 to 56.35 for the 120B model) and Burmese (4.98 to 31.35 for the 120B model), with instruction following, reasoning, and understanding improving most consistently.
Which Southeast Asian languages does SEA-LION-v4.8 support?
Evaluation covers seven languages: Burmese, Filipino, Indonesian, Malay, Tamil, Thai, and Vietnamese. The training data also includes Balinese, Javanese, Khmer, Lao, Sundanese, and Chinese through instruction and parallel corpora. Notably, Filipino receives no dedicated pre-training data — its improvement comes only from cross-lingual transfer — and the retained tokenizer is inefficient for Khmer, Lao, and Tamil scripts.
Are the SEA-LION-v4.8 models and training code publicly available?
The paper states that all four model checkpoints (base and post-trained at both scales) are released as open models, and evaluation scores are published on a live public leaderboard (leaderboard.sea-lion.ai). The SEA-Instruct training dataset is referenced by name. However, the paper does not explicitly state that the training or post-training code is released, and key details like the final weight-merging procedure are not specified.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-5 are described only via captions; the per-category breakdown behind Figure 1 and the exact tokenizer fertility values in Figures 3-4 could not be analyzed from extracted text.
- Tables 6-7 contain unexplained missing cells ('–' for Malay NLR, Tamil Knowledge, Burmese multi-turn), so capability comparisons for those language-task pairs are incomplete.
- Some inconsistencies (swapped 30B-A12B/120B-A3B designations, conflicting base-model names) appear verbatim in the extracted text and are treated as genuine paper errors, though extraction artifacts cannot be fully excluded.
- The paper references a live leaderboard for current scores; the review is based on the snapshot dated September 15, 2026 as stated in Table 5.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.