Rogue-27B-KR

Rogue-27B-KR — Korean-Specialized High-Performance Reasoning Model

<p align="center"> <img src="https://img.shields.io/badge/Korean_AI-Rogue--27B--KR-red?style=for-the-badge" alt="Rogue-27B-KR"> <img src="https://img.shields.io/badge/K--AI_Leaderboard-2nd_Place-gold?style=for-the-badge" alt="K-AI 2nd"> <img src="https://img.shields.io/badge/Base-Darwin--27B--Opus-blue?style=for-the-badge" alt="Base"> <img src="https://img.shields.io/badge/License-Apache_2.0-green?style=for-the-badge" alt="License"> </p>

Qwen3.5 Hybrid Architecture | ~26B Params | Thinking Mode | 262K Context | BF16 | Apache 2.0 Darwin-27B-Opus(Father) + Korean-SFT Qwen 27B(Mother) Merge + Korean SFT

---

K-AI Leaderboard — 2nd Place

K-AI Leaderboard is operated by the Korean Ministry of Science and ICT (MSIT) / NIA, evaluating Korean AI language models across multiple benchmarks.

Rogue-27B-KR achieved 2nd place on the K-AI Leaderboard with an average score of 0.549.

<p align="center"> <img src="assets/l1.png" width="800" alt="K-AI Leaderboard Ranking"> </p> <p align="center"> <img src="assets/l2.png" width="800" alt="K-AI Leaderboard Charts"> </p>

---

Model Overview

Rogue-27B-KR is a Korean-specialized high-performance reasoning model:

• Father: FINAL-Bench/Darwin-27B-Opus — Evolutionary merge model based on Qwen3.5-27B (inheriting Claude 4.6 Opus reasoning patterns)
• Mother: Korean-specialized SFT applied Qwen 27B variant
• Rogue: Merged both parents + additional Korean SFT

Combines Darwin's powerful reasoning with the mother model's Korean expressiveness, greatly enhancing Korean intelligence across culture, language, law, history, and more.

Key Features

• K-AI Leaderboard 2nd Place — Top-tier Korean AI performance verified by government evaluation
• Strong reasoning — Chain-of-Thought inherited from Darwin-27B-Opus
• Enhanced Korean cultural intelligence — +6%p improvement on CLIcK benchmark (0.794)
• 262K context — Ultra-long Korean document processing
• Thinking mode — Step-by-step reasoning via <think> tags
• BF16 format — Memory-efficient (~48GB), optimized for modern GPUs
• Apache 2.0 — Free for commercial use

---

Benchmark Results

CLIcK (Cultural and Linguistic Intelligence in Korean)

200 questions, 0-shot evaluation:

+6.07%p over Qwen3.5-27B, +5.40%p over Darwin-27B-Opus

Category Details

---

Model Specifications

---

VRAM Requirements

---

Usage

Requirements: transformers >= 4.57.0 (Qwen3.5 support). Latest recommended: pip install -U transformers

Transformers

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

tokenizer = AutoTokenizer.from_pretrained("ginigen-ai/Rogue-27B-KR")
model = AutoModelForCausalLM.from_pretrained(
    "ginigen-ai/Rogue-27B-KR",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [{"role": "user", "content": "Explain traditional Korean wedding ceremonies."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=4096, do_sample=False)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

vLLM

vllm serve ginigen-ai/Rogue-27B-KR \
    --enforce-eager \
    --max-model-len 32768 \
    --dtype bfloat16

SGLang

python -m sglang.launch_server \
    --model-path ginigen-ai/Rogue-27B-KR \
    --tp 1 \
    --mem-fraction-static 0.90 \
    --context-length 32768 \
    --dtype bfloat16 \
    --reasoning-parser qwen3

---

Training Info

Lineage

Qwen/Qwen3.5-27B (Apache 2.0)
+-- FINAL-Bench/Darwin-27B-Opus (evolutionary merge)  <- Father
|
+-- Korean-SFT Qwen 27B (Korean specialization)       <- Mother
        |
    [Merge: Father + Mother]
        |
    [Additional Korean SFT]
        |
    ginigen-ai/Rogue-27B-KR (this model, BF16)

---

Citation

@misc{ginigen_rogue_27b_kr_2026,
  title        = {Rogue-27B-KR: Korean-Specialized High-Performance Reasoning Model},
  author       = {GinigenAI},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/ginigen-ai/Rogue-27B-KR}}
}

---

Acknowledgements

• VIDRAFT — Darwin-27B-Opus base model
• Qwen Team — Qwen3.5 architecture
• Korean Ministry of Science and ICT / NIA — K-AI Leaderboard

---

Built by GinigenAI for the Korean AI community

학습 데이터셋

본 모델은 AI Hub(과학기술정보통신부·NIA 구축) 공개 한국어 말뭉치로 파인튜닝되었습니다.

✅ 전문 의학지식 데이터 ✅ 한국어 생성 기반 상식추론 데이터셋 ✅ 금융·법률 문서 기계독해 데이터 ✅ 국가기록물 초거대 AI 학습 말뭉치 ✅ 대규모 구매도서 기반 한국어 말뭉치 ✅ 멀티세션·대화 관련 말뭉치 데이터 ✅ 한국어 대화 데이터셋