Technical Report 👁️

Nemotron-SEA-LION-v4.8-120B-A12B-NVFP4

Last updated: 2026-09-18

This repository contains the NVFP4 (4-bit floating point, E2M1) quantized weights for aisingapore/Nemotron-SEA-LION-v4.8-120B-A12B.

The model is optimized for Southeast Asian (SEA) languages, cultures, and contexts. This repository provides quantized models that have significantly compressed memory footprints while preserving model accuracy, enabling low-latency MoE inference across high-throughput production environments.

NVFP4 is designed for NVIDIA Blackwell GPUs (B200/GB200), where it can run natively as W4A4 on FP4 tensor cores. On Hopper (H100/H200), vLLM serves it as W4A16 through the Marlin kernel, so the memory saving carries over but it is slower than the FP8 checkpoint.

Model Details

Model Description

SEA-LION stands for Southeast Asian Languages In One Network.

The model compression process quantizes only the routed MoE expert projections: the up_proj/down_proj of all 512 routed experts in each of the 40 MoE layers (20,480 routed-expert matrices). The weights are stored as 4-bit FP4 (E2M1) with a per-16-element FP8 block scale and an FP32 per-tensor scale, and input activations use a calibrated per-tensor scale for W4A4 kernels. The calibration was performed on 256 conversations from HuggingFaceH4/ultrachat_200k.

This model was quantized using NVIDIA Model Optimizer with the MAMBA_MOE_NVFP4_CONSERVATIVE_CFG preset, modified only to keep the shared expert in BF16.

For tokenization, the model employs the default tokenizer used in nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16.

• Developed by: AI Products Pillar, AI Singapore
• Funded by: National Research Foundation Singapore
• Shared by: AI Products Pillar, AI Singapore
• Model type: Instruction-tuned language model
• Architecture: Mamba2-Transformer Hybrid MoE
• Context length: 262,144 tokens
• Language(s): Burmese, English, Indonesian, Filipino, Malay, Tamil, Thai, and Vietnamese
• License: MIT
• Parent model: aisingapore/Nemotron-SEA-LION-v4.8-120B-A12B

Model Sources

• Collection: SEA-LION v4.8 - an aisingapore Collection

Available Quantized Versions

We provide multiple quantization formats to optimize deployment trade-offs between memory footprint and output quality.

• Nemotron-SEA-LION-v4.8-30B-A3B-NVFP4
• Nemotron-SEA-LION-v4.8-30B-A3B-FP8
• Nemotron-SEA-LION-v4.8-30B-A3B-GGUF
• Nemotron-SEA-LION-v4.8-120B-A12B-NVFP4
• Nemotron-SEA-LION-v4.8-120B-A12B-FP8
• Nemotron-SEA-LION-v4.8-120B-A12B-GGUF

Usage

vLLM

You can serve the model using vLLM:

from vllm import LLM, SamplingParams

llm = LLM(
    model="aisingapore/Nemotron-SEA-LION-v4.8-120B-A12B-NVFP4",
    trust_remote_code=True,
)

messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Where is Lau Pa Sat? Tell me in a single sentence."},
]

params = SamplingParams(temperature=0.7, top_p=0.9 )
outputs = llm.chat(messages, params)
print(outputs[0].outputs[0].text)

SGLang

You can serve the model using SGLang:

from sglang import Engine, SamplingParams

# SGLang natively detects the pre-quantized FP8 checkpoint
llm = Engine(
    model_path="aisingapore/Nemotron-SEA-LION-v4.8-120B-A12B-NVFP4",
    trust_remote_code=True,
)

messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Where is Lau Pa Sat? Tell me in a single sentence."},
]

# Configure generation constraints
params = SamplingParams(temperature=0.7, top_p=0.9)

# Execute inference using the OpenAI-compatible chat message format
outputs = llm.chat(messages, params)
print(outputs["text"])

Evaluation

Throughput Test

Each checkpoint was served with vLLM on NVIDIA H200 GPUs and benchmarked over the OpenAI-compatible chat endpoint with GuideLLM, an open-source throughput benchmark from the vLLM project.

The prompts are our own Southeast Asian instructions from the SEA-instruct dataset, streamed with a fixed output length per request so that every checkpoint generates the same number of tokens.

Time to first token is the time taken from sending the request until the first content token arrives, so it covers prompt processing (prefill) and network overhead on localhost. Tokens per second is a metric that measures the number of output tokens divided by that request's total time from send to final token. Both are reported as mean ± standard deviation over all requests completed in a 120-second run. The VRAM Usage is taken from vLLM upon loading.

Note: Benchmarks were captured on Hopper architecture (NVIDIA H200) GPUs; NVFP4 precision formats yield further hardware-level acceleration when executed natively on NVIDIA Blackwell infrastructure.

All models were served at their full context length of 262,144 tokens with default vLLM settings. The tokens per second are reported as the aggregate across the GPUs used.

FP8 checkpoints are a good default for reduced-precision deployment. NVFP4 checkpoints offer a 4-bit alternative designed for NVIDIA's Blackwell architecture.

The figures reflect Hopper execution, and the checkpoints have not yet been benchmarked on Blackwell. We recommend measuring on your own workload before choosing a format.

For details on Nemotron-SEA-LION-v4.8 performance, please refer to the SEA-HELM Leaderboard.

Technical Specifications

Technical Report

For training details, see the SEA-LION-v4.8 Technical Report.

Model Architecture

The architecture is based on the highly efficient Nemotron-3-Super foundation. The detailed architecture can be found at nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 documentation.

Uses

Out-of-Scope Use

The model has not been aligned for safety. Developers and users should perform their own safety fine-tuning and related security measures. In no event shall the authors be held liable for any claims, damages, or other liabilities arising from the use of the released weights and codes.

Bias, Risks, and Limitations

The model was not tested for robustness against adversarial prompting. It is important for users to be aware that our model exhibits certain limitations that warrant consideration. Like many LLMs, the model can hallucinate and occasionally generates irrelevant content, introducing fictional elements that are not grounded in the provided context. Users should also exercise caution in interpreting and validating the model's responses due to the potential inconsistencies.

Citation

BibTeX:

@misc{aisingapore2026sealionv48technicalreport,
      title={SEA-LION-v4.8: A Technical Report},
      author={Adila Aulia and Ahmed Dabeer and Ahn Jeongmi and Antonyrex Sajeban and Chan Hok Teng Adwin and Cheng Zi Yi Nicholas and Choa Hsueh Mei Esther and Heng Jonathan and Jann Railey Estrada Montalan and Lee Chwan Ren and Leong Wai Yi and Leong Wei Qi and Liew Rachel and Limkonchotiwat Peerat and Muhammad Ridzuan Bin Mokhtar and Nagarajan Karthik and Ng Boon Cheong Raymond and Ngee Chia Tai and Ngui Jian Gang and Nguyen Thanh Ngan and Ong Tat-Wee David and Pereira Mark and Phang Shi Wei Benjamin and Poon Joseph and Rengarajan Hamsawardhini and Susanto Yosephine and Sutaveephamochanon Anocha and Tan Choon Meng and Tan Chor Phin Evelyn and Tan Le Min Sheryl and Tan Siao Wei Jessica and Tan Yixian and Tasawong Panuthep and Tee Jun Yun and Teng Kok Wai Walter and Teo Eng Sipp Leslie and Tjhi William and Tuchinda Pume and Wu Donghang and Yong Xianbin and Zhang Zhou},
      year={2026},
      eprint={2609.18310},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2609.18310},
}

Team

AI Products Pillar, AI Singapore

Acknowledgement

This project is supported by the National Research Foundation Singapore and Infocomm Media Development Authority (IMDA), Singapore under its National Large Language Model Funding Initiative.

Contact

sealion@aisingapore.org