Announcing Qwen-SEA-LION-v4: Advanced Reasoning and Language Depth for Southeast Asia

ByAISG | AI Products

We are proud to introduce Qwen-SEA-LION-v4-32B-IT, a new iteration of our powerful multilingual model series specifically trained for Southeast Asia. This initial experimental version represents a significant architectural evolution, transitioning from a sentence-piece tokenizer to a more modern byte-pair encoding (BPE) for superior text handling.

Built upon the powerful Qwen3-32B foundation model, Qwen-SEA-LION-v4-32B-IT inherits long-context capabilities, broad multilingual support, and modern developer features like function calling, while deepening its specialization in the languages, cultures, and contexts of Southeast Asia.

What’s New in Qwen-SEA-LION-v4?

This release focuses on three core areas of improvement:

  1. Significantly Enhanced Capabilities in SEA Languages: The model’s enhanced capabilities stem from a two-pronged approach: inheriting the broad power of the Qwen3 architecture and deepening it with focused SEA-specific training. From Qwen3, it gains support for over 100 languages and high performance in reasoning, math, and code. Our post-training then adds a crucial layer of specialization, improving contextual fluency with demonstrated gains in Burmese, Filipino, Indonesian, Malay, Tamil, Thai, and Vietnamese.
  2. Developer-Friendly Features: Integration with the Qwen3 family provides access to modern tools and features, facilitating easier adoption and the development of more sophisticated applications.
  3. Efficient Performance with Minimal Trade-offs: To make our state-of-the-art model more accessible, we are also releasing 4-bit and 8-bit quantized versions. These models dramatically reduce memory and computational requirements, allowing them to run on a wider range of hardware. This efficiency comes at a remarkably low cost: no performance degradation for the 8-bit version and a 0.3% drop in performance for the 4-bit version when compared to the full 16-bit version, giving you near full-precision performance with significantly lower deployment costs.

Model Specifications and Availability

AttributeQwen-SEA-LION-v4-32B-IT
Base ArchitectureQwen3-32B
Parameters31.2B
Context Length32k tokens natively


The Qwen-SEA-LION v4 models are available in the following versions:

Technical Deep Dive: Training and Evaluation

Training Data

The model’s specialized capabilities are derived from a comprehensive pretraining dataset totaling 100 billion tokens. This dataset, which is a subset of SEA-PILE-v2, includes a rich mixture of sources covering Burmese, Indonesia, Malay, Tagalog, Tamil, Thai, and Vietnamese. The data was collected from web crawls, source code, open-source academic datasets, and high-quality synthetically generated content.

Training Procedure

The model underwent continued pre-training on the 7 SEA languages. Our post-training workflow is a multi-stage process designed to maximize performance and alignment. We utilized Instruction Fine-Tuning to refine the base model. The initial stage used a combined dataset including our own SEA-Instruct, Infinity-Instruct, and OpenMath-Instruct 2, alongside robust open-source datasets.

Evaluation Framework

Qwen-SEA-LION v4 was rigorously evaluated using a dual-pronged approach to measure both its general and region-specific capabilities.

  • English Evaluation: We used a standard suite of challenging English benchmarks, including BBH, GPQA, IFEval, Math-Hard, MMLU-Pro, and MUSR.
  • Southeast Asian Evaluation: Performance was measured using SEA-HELM, our holistic benchmark covering Burmese, Filipino, Indonesian, Malay, Tamil, Thai, and Vietnamese. Tasks included summarization, toxicity detection, instruction following (SEA-IFEval), and conversational ability (SEA-MTBench).

All evaluations were conducted in a zero-shot setting using native language prompts.

Performance and Leaderboard Standing

The results underscore the effectiveness of our specialized approach. Based on the comprehensive SEA-HELM evaluation suite:

Qwen-SEA-LION v4 ranks #6 out of 59 models on the SEA-HELM benchmark and, crucially, #1 among all open models under 200B parameters.

This remarkable performance demonstrates that a regionally specialized model can outperform much larger, more generalized systems on relevant tasks. Furthermore, it achieves this while being efficient enough to run on a consumer-grade laptop with 35GB of RAM.

For detailed results and comparisons, please visit our official leaderboard: https://sea-lion.ai/leaderboard

AI Singapore is a national programme supported by the National Research Foundation, Singapore and hosted by the National University of Singapore. Any opinion, finding, conclusion or recommendation expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore, or the National University of Singapore.

We also grateful for the support of the Infocomm Media Development Authority (IMDA) of Singapore.