websfactory/Webs-Sejong-31B-v7

VISIONConcurrent Unit Cost:2Model Size:31BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 3, 2026License:gemmaArchitecture:Transformer0.0K Featherless Exclusive Cold

Webs-Sejong-31B-v7 by websfactory is a 31-billion parameter Korean-centric language model based on Google's Gemma-4-31B-it architecture, featuring a 32768 token context length. It excels in Korean-language knowledge, cultural context, and academic reasoning, while maintaining English capabilities. This model achieved the #1 rank on the K-AI Leaderboard with an overall average of 0.624, making it ideal for applications requiring strong Korean linguistic and cultural understanding.

Loading preview...

Webs-Sejong-31B-v7: A Korean-Centric Gemma-4 Model

Webs-Sejong-31B-v7 is a 31-billion parameter language model developed by websfactory, built upon the google/gemma-4-31B-it architecture. It is specifically optimized for Korean language tasks and cultural understanding, while also retaining strong English capabilities.

Key Capabilities & Highlights

  • #1 on K-AI Leaderboard: Achieved the top rank on the public K-AI Leaderboard (as of 2026-07-05) with an overall average score of 0.624, evaluated on non-public benchmark data.
  • Korean-First Design: Demonstrates strong performance in Korean-language knowledge, cultural context, professional and academic reasoning, and commonsense QA.
  • Gemma-4 Compatibility: Utilizes the standard Gemma-4 architecture and tokenizer, ensuring compatibility with Hugging Face transformers Gemma-4 implementations and existing serving stacks.
  • Architecture: Dense Gemma-4-31B with bfloat16 precision.

Intended Use Cases

This model is primarily intended for applications requiring robust Korean-language assistance, knowledge question-answering, and reasoning. Its strong performance on the K-AI Leaderboard suggests its suitability for tasks demanding high accuracy in a Korean context. Users should evaluate its performance on their specific target tasks before production deployment.