websfactory/Webs-Sejong-31B-v2
Webs-Sejong-31B-v2 is a 31-billion parameter Korean-centric language model from websfactory, built on the Gemma-4 architecture through weight-space model merging. Optimized for Korean cultural knowledge and academic/professional reasoning, it retains English ability from its base model. This model achieved 2nd place on the NIA K-AI Leaderboard, demonstrating strong reasoning capabilities in Korean benchmarks.
Loading preview...
Webs-Sejong-31B-v2: Korean-Centric Gemma-4 Merge Model
Webs-Sejong-31B-v2 is a 31-billion parameter language model developed by websfactory, leveraging the Gemma-4 architecture through a weight-space model merging technique. It is explicitly stated as a merge model, not a separately trained one, with no additional training applied. The model is specifically optimized for Korean cultural knowledge and academic/professional reasoning, while maintaining the English capabilities inherited from its base architecture.
Key Capabilities & Performance
- Top-tier Korean Performance: Achieved 2nd place on the NIA K-AI Leaderboard with an average score of 0.620 on a private Korean benchmark suite (KMMLU-Pro, CLIcK, HLE, MuSR, Com2), just 0.001 behind the 1st-place model.
- Strong Reasoning: Outperforms the 1st-place model on the MuSR (Korean reasoning) benchmark (0.617 vs 0.591) and HLE (Korean).
- Korean-First Design: Tuned for deep understanding of Korean culture and professional/academic contexts.
- Standard Integration: Utilizes the standard Gemma-4 architecture and tokenizer, allowing for seamless loading in
transformersand serving with vLLM without custom code.
Intended Use Cases
- Korean-language assistance: Ideal for applications requiring nuanced understanding and generation in Korean.
- Knowledge QA: Excels in question-answering tasks related to Korean cultural, academic, and professional domains.
- Reasoning tasks: Particularly strong in complex reasoning challenges within the Korean language context.