Qwen-SEA-LION-v4.5-27B-IT — FP8 Dynamic

FP8 (dynamic) quantization of aisingapore/Qwen-SEA-LION-v4.5-27B-IT, produced with llmcompressor using the FP8_DYNAMIC scheme.

• Quantization: per-tensor dynamic FP8 (w8a8), no calibration data needed
• Size: ~29.5 GB (vs 54.5 GB BF16)
• Inference: vLLM ≥ 0.7 with --quantization fp8
• Quality: dynamic FP8 has negligible accuracy degradation vs BF16 on standard benchmarks

About the base model

Qwen-SEA-LION-v4.5-27B-IT built upon the Qwen3.6-27B dense architecture, a 27-billion parameter model featuring a hybrid Linear and Full Attention design. To ensure deep domain adaptation, the model underwent distillation from Qwen3.5-397B-A17B on an updated aisingapore/SEA-Instruct-2602, instilling multilingual and multicultural fluency across English and key SEA languages including: Burmese, Indonesian, Filipino, Malay, Tamil, Thai, and Vietnamese.

Key features inherited from Qwen3.6:

• Context Window (262K): Large context window to enable strong reasoning capabilities.
• Reasoning: Highly capable reasoning model, with configurable thinking modes.
• Thinking Preservation: Retains historical reasoning context to streamline iterative development and reduce compute overhead.
• Agentic Coding: High-precision handling of repository-level reasoning and frontend workflows.
• Unified Vision-Language: Early fusion training delivers good performance across multimodal reasoning, coding, and visual tasks.

Quickstart (vLLM)

pip install vllm>=0.7

vllm serve toeniant/Qwen-SEA-LION-v4.5-27B-IT-FP8-Dynamic \
  --quantization fp8 \
  --max-model-len 262144

from vllm import LLM, SamplingParams

llm = LLM(
    model="toeniant/Qwen-SEA-LION-v4.5-27B-IT-FP8-Dynamic",
    quantization="fp8",
    max_model_len=262144,
)

sampling = SamplingParams(
    temperature=0.7,
    top_p=0.80,
    top_k=20,
    max_tokens=512,
)

messages = [
    {"role": "user", "content": "Tolong carikan flat 4-bilik dekat Tampines, bajet bawah $500,000."}
]

out = llm.chat(messages, sampling_params=sampling)
print(out[0].outputs[0].text)

Model Details

• Developed by: AI Products Pillar, AI Singapore (base model)
• Quantized by: toeniant
• Model type: Causal Language Model with Vision Encoder
• Context length: 262k
• Language(s): English, Burmese, Indonesian, Filipino, Malay, Tamil, Thai, Vietnamese
• License: MIT
• Base model: aisingapore/Qwen-SEA-LION-v4.5-27B-IT

Evaluation

For benchmark results, please refer to the SEA-LION Leaderboard and the original model card.

Bias, Risks, and Limitations

The model has not been aligned for safety. Developers and users should perform their own safety fine-tuning and related security measures. Like many LLMs, the model can hallucinate and occasionally generates irrelevant content. Users should exercise caution in interpreting and validating the model's responses.

Acknowledgement

Original model by AI Singapore, supported by the National Research Foundation Singapore and IMDA under its National Large Language Model Funding Initiative. See the original model card for full credits and team information.

Contact

Original authors: sealion@aisingapore.org