Nemotron-SEA-LION-v4.8-120B-A12B
Nemotron-SEA-LION-v4.8-120B-A12B is the larger v4.8 instruction model. It starts from AI Singapore's 120B-A12B base release and uses supervised fine-tuning with online on-policy distillation. Its Mamba2-Transformer hybrid design uses mixture-of-experts routing.
How this release works
The card lists English and seven regional languages: Burmese, Indonesian, Filipino, Malay, Tamil, Thai and Vietnamese. Training uses a subset of SEA-Instruct-2602. The release uses the tokenizer associated with NVIDIA Nemotron 3 Super 120B-A12B.Nemotron-SEA-LION-v4.8-120B-A12B official model card
How to assess it for your work
Practical guidance. Use this model in a comparison where the smaller release misses relationships across several passages or fails a multi-step instruction. Judge the resulting answer against the supplied evidence. A larger model is useful only when its additional resource demand produces a measurable improvement in your own task.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | aisingapore/Nemotron-SEA-LION-v4.8-120B-A12B |
| Context window | 262,144 tokens Context length stated for the released instruction-tuned checkpoint. Nemotron-SEA-LION-v4.8-120B-A12B official model card |
| Model family size | 120B-A12BNemotron-SEA-LION-v4.8-120B-A12B official model card |
| Parent checkpoint | Nemotron-SEA-LION-v4.8-120B-A12B-BaseNemotron-SEA-LION-v4.8-120B-A12B official model card |
| License declared by the model card | MITNemotron-SEA-LION-v4.8-120B-A12B official model card |
| Input and output | TextNemotron-SEA-LION-v4.8-120B-A12B official model card |
Related models
- Nemotron-SEA-LION-v4.8-30B-A3B
AI Singapore ยท SEA-LION v4.8
A regional instruction model that combines Mamba2 and Transformer layers with mixture-of-experts routing.
Launch date not confirmedDownloadable