websfactory/Webs-Sejong-31B-v3
Webs-Sejong-31B-v3 is a 31-billion parameter Korean-centric language model developed by websfactory, built on the Gemma-4 architecture through weight-space model merging. This model is optimized for Korean cultural knowledge, commonsense, and academic/professional reasoning, while retaining English language ability. It is designed to handle multi-step reasoning tasks and is a drop-in replacement for standard Gemma-4 models, loading directly in `transformers`.
Loading preview...
Overview
Webs-Sejong-31B-v3 is a 31-billion parameter language model from websfactory, created by merging models based on the Gemma-4 architecture. It is explicitly stated as a merge model, not a separately trained one, with no additional training applied. The model retains the standard Gemma-4 architecture and tokenizer, ensuring compatibility with existing transformers and vLLM setups without custom code.
Key Capabilities
- Korean-first Optimization: Specifically tuned for Korean cultural knowledge, commonsense, and professional/academic reasoning.
- Reasoning-Oriented: Designed to excel in multi-step reasoning tasks while maintaining strong Korean language abilities.
- English Retention: Inherits and preserves English language capabilities from its base model.
- Ease of Use: Functions as a drop-in replacement, loading seamlessly with
transformersand serving in vLLM.
Intended Use & Limitations
This model is intended for Korean-language assistance, knowledge question-answering, and reasoning tasks. As a merged model, it inherits characteristics and potential biases from its source models, necessitating evaluation for specific use cases before deployment. It is distributed under the Gemma Terms of Use.