SEA-LION is a collection of Large Language Models (LLMs) which have been pretrained and instruct-tuned
for the Southeast Asia (SEA) region.
As of 25 Aug 2025, Gemma-SEA-LION-v4-27B-IT excels at Southeast Asian (SEA) tasks when compared to other open models
with fewer than 200 billion parameters and demonstrates performance comparable to that of larger and top closed models.
Gemma-SEA-LION-v4-27B-IT was quantized to create Gemma-SEA-LION-v4-27B-IT-NVFP4.
Gemma-SEA-LION-v4-27B-IT-NVFP4 has little degradation (<0.5%) in performance compared to Gemma-SEA-LION-v4-27B-IT
(for detailed rankings, please refer to the leaderboard), and
this version can run on a laptop with vLLM and CUDA-compatible GPU.
Image and text understanding capabilities, including document comprehension, visual Q&A, and image-grounded reasoning
Advanced function calling and structured outputs to allow for seamless integration into larger systems
Model Description
SEA-LION stands for Southeast Asian Languages In One Network.
Quantization was performed on Gemma-SEA-LION-v4-27B-IT to produce optimized variants that reduce memory requirements
while maintaining model quality. These quantized models support inference on a range of consumer-grade GPUs
and are compatible with various inference engines.
For tokenization, the model employs the default tokenizer used in Gemma 3 27B Instruct.
The dataset comprises Burmese, English, Indonesian, Khmer, Lao, Malay, Mandarin, Tagalog, Tamil, Thai and Vietnamese languages, collected from a mixture of sources including web data, code, open-source datasets, and synthetically generated datasets, amounting to a total of 500 billion tokens sampled from our bucket of 1 trillion tokens.
Evaluation
Testing Data, Factors & Metrics
Testing Data
We evaluated Gemma-SEA-LION-v4-27B-IT-NVFP4 on general language, multi-turn chat and instruction-following capabilities.
Testing Data
General language capabilities
For the evaluation of general language capabilities, we employed the SEA-HELM evaluation benchmark across a variety of tasks.
These tasks include Question Answering (QA), Sentiment Analysis (Sentiment), Toxicity Detection (Toxicity), Translation in both directions (Eng>Lang & Lang>Eng),
Abstractive Summarisation (Abssum), Causal Reasoning (Causal), Natural Language Inference (NLI), Linguistic Diagnostics (LINDSEA), Cultural Knowledge (Kalahi)
and Global MMLU Lite.
Instruction-following and Multi-turn Chat
We evaluated the models on instruction-following and multi-turn chat capabilities with SEA-IFEval (based on IFEval) and SEA-MTBench (based on MT-Bench) respectively.
The two datasets were originally in English, the linguists and native speakers in the team worked together to filter, localise and translate the datasets into the respective target languages to ensure that the examples remained reasonable, meaningful and natural.
Factors
All evaluations were run with the model specific generation parameters defined in the model config. Each evaluation comprised of 8 runs with different seeds and the final results were averaged across these runs.
For all tasks, the model was expected to provide an answer tag from which the answer was automatically extracted. For tasks where options were provided, the answer should comprise one of the pre-defined options.
The evaluation was done zero-shot with native prompts on a sample of 100-1000 instances for each dataset.
SEA-IFEval
SEA-IFEval evaluates a model's ability to adhere to constraints provided in the prompt,
for example beginning a response with a specific word/phrase or answering with a certain number of sections.
Additionally, accuracy is normalised by the proportion of responses in the correct language
(if the model performs the task correctly but responds in the wrong language, it is judged to have failed the task).
SEA-MTBench
SEA-MTBench evaluates a model's ability to engage in multi-turn (2 turns) conversations and respond in ways that align with human needs.
We use gpt-4.1-2025-04-14 as the judge model and compare against gpt-4.1-2025-04-14 as the baseline model.
The metric used is the weighted win rate against the baseline model (i.e. average win rate across each category:
Math, Reasoning, STEM, Humanities, Roleplay, Writing, Extraction).
Metrics
The following metrics were used:
Task
Metric
Sentiment Analysis
Accuracy
Extractive QA (ID, VI, TH, TA)
ChrF++
MCQ-QA (TL, MY, MS)
Accuracy
Metaphor
Accuracy
Abstractive Summarisation
Rouge-L
Translations
MetricX-24 score (with reference)
Causal Reasoning
Accuracy
Natural Language Inference
Accuracy
LINDSEA
Accuracy
Global MMLU Lite
Accuracy
Kalahi
Accuracy
SEA-IFEval
Accuracy
SEA-MTBench
Win rate against a reference
Toxicity Detection
Accuracy
Results
Leaderboard results on SEA-HELM snapedshot on 25 Aug 2025!
For details on Gemma-SEA-LION-v4-27B-IT-NVFP4 performance, please refer to the SEA-HELM leaderboard, Leaderboard results on SEA-HELM.
Additional remarks:
SEA-HELM evaluation suite run on one single H200
Default model parameters as listed in the model card used for evaluation
Models served through vLLM
Model Performance Across Different GPU Architectures
GPU Chip
Model Variant
Actual VRAM Usage (GB)
Time to First Token (s)
Tokens per Second
GPU Utilization (%)
Notes
H100
IT (Full Precision)
51.4 GB
0.0457 ± 0.0016
28.57 ± 0.08
95.3 ± 1.6
Baseline full precision
H100
IT-FP8-Dynamic
27.6 GB
0.0338 ± 0.0028
46.87 ± 0.23
93.3 ± 0.5
FP8 quantization
H100
IT-NVFP4
17.3 GB
0.0276 ± 0.0019
62.43 ± 0.08
90.5 ± 1.1
FP4 quantization
A100
IT (Full Precision)
51.4 GB
0.0503 ± 0.0013
25.63 ± 0.03
96.8 ± 0.4
Baseline full precision
A100
IT-FP8-Dynamic
27.6 GB
0.0334 ± 0.0020
41.80 ± 0.11
95.0 ± 0.0
FP8 quantization
A100
IT-NVFP4
17.3 GB
0.0300 ± 0.0026
56.28 ± 0.23
94.0 ± 0.0
FP4 quantization
L40S
IT-FP8-Dynamic
27.6 GB
0.0549 ± 0.0012
21.26 ± 0.00
96.9 ± 0.3
FP8 quantization
L40S
IT-NVFP4
17.3 GB
0.0357 ± 0.0013
35.81 ± 0.01
95.0 ± 0.0
FP4 quantization
Additional Remarks:
TTFT and Tokens per Second: measured with vLLM on localhost and concurrency = 1.
Reported results are the median (p50) values, calculated across 10 requests.
Model size taken from vLLM upon loading
Tests conducted on a system with a single GPU Chip.
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.
Limitations
In terms of vision capability, Gemma-SEA-LION-v4-27B-IT has been trained and fine-tuned exclusively on the text back-end.
As a result, its vision capabilities are expected to be comparable to those of Gemma 3 IT 27B,
and may not exhibit significant improvements or differences in this area. 🤗 google/gemma-3-27b-it
More Information
This is the repository for the commercial instruction-tuned model.
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.
AI Singapore is a national programme supported by the National Research Foundation, Singapore and hosted by the National University of Singapore.
Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of the National Research Foundation or the National University of Singapore.
Antonyrex Sajeban, Chan Hok Teng Adwin, Cheng Zi Yi Nicholas, Choa Hsueh Mei Esther, Heng Jonathan, Huang Yuli, Hulagadri Adithya Venkatadri, Jann Railey Estrada Montalan,
Kang Siow Wei Bryan, Lau Wayne, Lee Chwan Ren, Leong Wai Yi, Leong Wei Qi, Limkonchotiwat Peerat, Muhammad Ridzuan Bin Mokhtar, Nagarajan Karthik, Ng Boon Cheong Raymond,
Ngee Chia Tai, Ngui Jian Gang, Nguyen Thanh Ngan, Ong Jin Jie Brandon, Ong Tat-Wee David, Ong Zhi Hao, Pereira Mark, Rengarajan Hamsawardhini, Susanto Yosephine,
Sutaveephamochanon Anocha, Tan Choon Meng, Tan Chor Phin Evelyn, Tan Siao Wei Jessica, Teng Kok Wai Walter, Teo Eng Sipp Leslie, Tjhi William, Yeo Yeow Tong, Yong Xianbin,
Liew Rachel, Liu Bing Jie Darius, Teo Wei Yi, Zhou Lin (NCS), Gopalakrishnan Roshan (NCS), Anda Cuahtemoc (NCS), Sri Devi Wijaya (NCS), Nandi Partha (NCS),
Elliott Chris (Google), Mohseni Mohammadreza (Google), Sharan Mayank (Google), Wei Fanny (Google), Tang Jiuqiang (Google), Xu Xiang (Google), Yu Ting (Google),
Loh Michelle (Google), Mangal Saurabh (Google), Mukherjee Pratyusha (Google), Sim Stephanie (Google)
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.
