---
license: apache-2.0
language:
- ko
library_name: transformers
pipeline_tag: image-text-to-text
base_model: VIDraft/JGOS-31B-Citizen
tags:
- korean
- reasoning
- think
- gemma4
---

# JGOS-31B-Think

Korean reasoning-focused LLM (31B, Gemma4) from VIDRAFT. Native step-by-step think reasoning.

## Docker Deployment (K-AI Evaluation)

Image: vidraft/jgos-31b-think:01.03  
vLLM 0.22.0 (model baked-in), Port 8000 (OpenAI-compatible API).

Memory-tuned for single eval GPU: max-model-len 8192, vision disabled (text eval), max-num-seqs 16, gpu-memory-utilization 0.90. bf16 weights ~59GB (needs a single GPU >= ~64GB; for smaller GPUs use FP8).

## License
Gemma license (inherited from base).