SEA-LION

weights

open(safetensors on HF for all four v4.8 checkpoints, ungated)

data

partial(the supervised examples are a subset of the published aisingapore/SEA-Instruct-2602 (ODC-BY, automatic click-through)

code

partial(usage code only

license

MIT(OSI

SEA-LION v4.8, the current release, declares MIT on every checkpoint, but the models are built on NVIDIA's Nemotron 3 and ship with its Nemotron Open Model License, which allows commercial use without being an open-source license. Post-training drew on a published Southeast Asian instruction dataset, though only a subset of it, and the training pipeline is described rather than released. The Gemma-based v4 model, under Google's Gemma terms, belongs to an older release.

• https://arxiv.org/html/2609.18310 recorded 2026-10-13

  SEA-LION-v4.8 technical report. Offline supervised examples "are a subset from our aisingapore/SEA-Instruct-2602 dataset"; a second source of post-training data "consists of trajectories generated dynamically by the model". Section 4.2 describes a five-part post-training pipeline; searched the body for a code repository link and found none.

• https://huggingface.co/aisingapore/Nemotron-SEA-LION-v4.8-120B-A12B/raw/main/NOTICE.md recorded 2026-10-13

  Third-party notices. "AI Singapore opensource models are under NTU-mandated MIT License", followed by a "Base Model License (nvidia-nemotron-open-model)" section reproducing the NVIDIA Nemotron Open Model License of December 15, 2025: "Works are commercially usable", "You are free to create and distribute Derivative Works", and redistribution requires giving recipients "a copy of this License". It names no user or revenue ceiling.

• https://huggingface.co/aisingapore/Nemotron-SEA-LION-v4.8-120B-A12B/raw/main/README.md recorded 2026-10-13

  Card: built on Nemotron-SEA-LION-v4.8-120B-A12B-Base "by fine-tuning with a mixture of SFT and online on-policy distillation (OPD)"; "License: MIT"; under Training Data, "The training data is a subset of aisingapore/SEA-Instruct-2602". The code on the card is a transformers usage example; no training code is linked.

• https://huggingface.co/api/collections/aisingapore/sea-lion-v48 recorded 2026-10-13

  The "SEA-LION v4.8" collection lists the 120B-A12B and 30B-A3B post-trained and base checkpoints with their FP8, NVFP4 and GGUF variants, every one ungated, and the technical report 2609.18310.

• https://huggingface.co/api/datasets/aisingapore/SEA-Instruct-2602 recorded 2026-10-13

  Dataset record: "gated": "auto", license "odc-by", parquet files per language. Its description calls it "a preliminary release of instruction-tuning data" that "contains only the filtered subset of data".

• https://huggingface.co/api/models?author=aisingapore&limit=200 recorded 2026-10-13

  AI Singapore's model listing. Every Nemotron-SEA-LION-v4.8 repository, created between 2026-09-03 and 2026-09-16, carries the tag license:mit. The Gemma-SEA-LION-v4-27B family, created in August 2025, carries license:gemma, and Qwen-SEA-LION-v4-32B-IT, created in October 2025, license:mit.

• https://huggingface.co/api/models/aisingapore/Gemma-SEA-LION-v4-27B-IT recorded 2026-10-13

  Hub record for the v4 Gemma model: "gated": false, cardData license "gemma", base_model aisingapore/Gemma-SEA-LION-v4-27B, and new_version "aisingapore/Qwen-SEA-LION-v4.5-27B-IT", so a later release supersedes it.

• https://huggingface.co/api/models/aisingapore/Nemotron-SEA-LION-v4.8-120B-A12B recorded 2026-10-13

  Hub record: "private": false, "gated": false, cardData license "mit", datasets ["aisingapore/SEA-Instruct-2602"], base_model aisingapore/Nemotron-SEA-LION-v4.8-120B-A12B-Base; files include model-00001-of-00050.safetensors onward, NOTICE.md and NVIDIA-Nemotron-Open-Model-License-12-12-25.pdf.

Adoption is counted as Hugging Face downloads summed across the three declared chat models: the current v4.8 120B model and the earlier v4 models built on Qwen3 and Gemma 3. AI Singapore publishes quantized and smaller checkpoints separately, and those are not counted.

• https://huggingface.co/api/models/aisingapore/Gemma-SEA-LION-v4-27B-IT recorded 2026-10-13

  "downloads":1066 over the trailing 30 days for aisingapore/Gemma-SEA-LION-v4-27B-IT.

• https://huggingface.co/api/models/aisingapore/Nemotron-SEA-LION-v4.8-120B-A12B recorded 2026-10-13

  "downloads":1592 over the trailing 30 days for aisingapore/Nemotron-SEA-LION-v4.8-120B-A12B.

• https://huggingface.co/api/models/aisingapore/Qwen-SEA-LION-v4-32B-IT recorded 2026-10-13

  "downloads":833 over the trailing 30 days for aisingapore/Qwen-SEA-LION-v4-32B-IT.

No general benchmark or arena reading of SEA-LION itself exists yet: Artificial Analysis does not list it, and a score read off its Nemotron base would describe a different model. Its own technical report measures it on SEA-HELM, the Southeast Asian benchmark suite, but no other model on the map has a SEA-HELM reading to place it against. It stays unrated until SEA-LION has a direct general reading or a same-benchmark comparison.

• https://artificialanalysis.ai/leaderboards/models recorded 2026-10-13

  The leaderboard's embedded model records give intelligenceIndex 14.19 for Mistral Medium 3.5, 12.83 for Nemotron 3 Super 120B A12B (Reasoning), 11.27 for Mistral Small 4 (Reasoning), 8.99 for Mistral Small 4 (Non-reasoning) and 8.44 for GPT-4o (Nov '24). No SEA-LION or Bielik model appears on the page. The digest drifts between fetches because the page carries live pricing and throughput figures.

• https://arxiv.org/abs/2609.18310 recorded 2026-10-13

  "SEA-LION-v4.8: A Technical Report", submitted 16 Sep 2026. Abstract: the family is "built upon NVIDIA Nemotron 3"; "On SEA-HELM, the 30B-A3B model improves the overall SEA score from 46.06 to 51.57, while the 120B-A12B model improves from 49.30 to 63.44."