aisingapore/Qwen-SEA-LION-v4.5-27B-IT

VISIONConcurrency Cost:2Model Size:27BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:May 14, 2026License:mitArchitecture:Transformer0.0K Open Weights Cold

Qwen-SEA-LION-v4.5-27B-IT is a 27-billion parameter causal language model developed by AI Singapore, built upon the Qwen3.6 dense architecture with a hybrid Linear and Full Attention design. It has been extensively post-trained and instruct-tuned for the Southeast Asia (SEA) region, demonstrating multilingual and multicultural fluency across English and key SEA languages. This model excels in reasoning, agentic coding, and unified vision-language tasks, featuring a large 262K context window and retaining historical reasoning context for iterative development.

Loading preview...

Model Overview

AI Singapore's Qwen-SEA-LION-v4.5-27B-IT is a 27-billion parameter causal language model, part of the SEA-LION collection, specifically designed and optimized for the Southeast Asia (SEA) region. Built on the Qwen3.6 dense architecture, it features a hybrid Linear and Full Attention design and a substantial 262K context window. The model underwent distillation from Qwen3.5-397B-A17B on an updated aisingapore/SEA-Instruct-2602 dataset, enhancing its multilingual and multicultural fluency across English, Burmese, Indonesian, Filipino, Malay, Tamil, Thai, and Vietnamese.

Key Capabilities

  • Multilingual Fluency: Proficient in English and seven key SEA languages, achieved through targeted post-training.
  • Advanced Reasoning: Inherits strong reasoning capabilities from the Qwen3.6 foundation, with configurable thinking modes.
  • Agentic Coding: High-precision handling of repository-level reasoning and frontend workflows.
  • Unified Vision-Language: Early fusion training provides good performance across multimodal reasoning, coding, and visual tasks.
  • Context Preservation: Retains historical reasoning context, streamlining iterative development and reducing compute overhead.

Use Cases

This model is particularly well-suited for applications requiring deep understanding and generation in Southeast Asian languages, complex reasoning tasks, and agentic workflows. It has been evaluated on SEA-HELM for general language capabilities and SEA-IFEval/SEA-MTBench for instruction-following and multi-turn chat, demonstrating robust performance in these areas. Developers should note that the model has not been aligned for safety and requires further safety fine-tuning for production use.