Views
No views yet
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 is operated by the Korean Ministry of Science and ICT (MSIT) / NIA, evaluating Korean AI language models across multiple benchmarks.
| Metric | Score |
|---|---|
| Average | 0.549 |
| KMMLU-Pro | 0.658 |
| CLIcK | 0.794 |
| HLE(Ko) | 0.07 |
| MuSR(Ko) | 0.584 |
| Com2-main(Ko) | 0.646 |
<think> tags| Model | CLIcK (Overall) | Culture | Language |
|---|---|---|---|
| Qwen3.5-27B (base) | 69.52% | 71.84% | 64.66% |
| Darwin-27B-Opus | 70.19% | 72.91% | 64.47% |
| Rogue-27B-KR | 75.59% | 77.85% | 70.86% |
+6.07%p over Qwen3.5-27B, +5.40%p over Darwin-27B-Opus
| Category | Qwen3.5-27B | Darwin-27B-Opus | Rogue-27B-KR |
|---|---|---|---|
| Economy | 93.22% | 93.22% | 94.92% |
| Geography | 70.23% | 70.23% | 75.57% |
| History | 47.00% | 47.00% | 53.50% |
| K-pop | 92.68% | 97.56% | 92.68% |
| Law | 59.50% | 60.00% | 69.50% |
| Politics | 80.95% | 82.14% | 85.71% |
| Society | 87.00% | 89.00% | 90.00% |
| Tradition | 81.50% | 82.50% | 88.50% |
| Function Words | 68.18% | 67.42% | 75.00% |
| Grammar | 44.50% | 44.50% | 53.00% |
| Text | 82.50% | 82.50% | 86.00% |
| Property | Value |
|---|---|
| Architecture | Qwen3.5 (GatedDeltaNet Hybrid Attention, 64-layer) |
| Parameters | ~26B (text-only, no vision encoder) |
| Hidden Size | 5120 |
| Intermediate Size | 16384 |
| Layers | 64 |
| Attention Heads | 24 (GQA, KV Heads: 4) |
| Context Length | 262,144 tokens (1M with YaRN) |
| Precision | BF16 (~48GB) |
| Vocab Size | 248,320 |
| Thinking | Supported (<think> tags) |
| Languages | Korean, English, Japanese, Chinese + 201 languages |
| License | Apache 2.0 |
| Setup | VRAM | Notes |
|---|---|---|
| BF16 (native) | ~48 GB | Single H100/B200 or 2x A100 |
| 4-bit quantized | ~14 GB | Single RTX 4090 |
| 8-bit quantized | ~26 GB | Single A6000 |
Requirements:transformers >= 4.57.0(Qwen3.5 support). Latest recommended:pip install -U transformers
1from transformers import AutoTokenizer, AutoModelForCausalLM
2import torch
3
4tokenizer = AutoTokenizer.from_pretrained("ginigen-ai/Rogue-27B-KR")
5model = AutoModelForCausalLM.from_pretrained(
6 "ginigen-ai/Rogue-27B-KR",
7 torch_dtype=torch.bfloat16,
8 device_map="auto",
9)
10
11messages = [{"role": "user", "content": "Explain traditional Korean wedding ceremonies."}]
12text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
13inputs = tokenizer(text, return_tensors="pt").to(model.device)
14outputs = model.generate(**inputs, max_new_tokens=4096, do_sample=False)
15print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))1vllm serve ginigen-ai/Rogue-27B-KR \
2 --enforce-eager \
3 --max-model-len 32768 \
4 --dtype bfloat161python -m sglang.launch_server \
2 --model-path ginigen-ai/Rogue-27B-KR \
3 --tp 1 \
4 --mem-fraction-static 0.90 \
5 --context-length 32768 \
6 --dtype bfloat16 \
7 --reasoning-parser qwen3| Item | Details |
|---|---|
| Father model | FINAL-Bench/Darwin-27B-Opus — Reasoning |
| Mother model | Korean-SFT Qwen 27B variant — Korean expressiveness |
| Merge | Weight merge (intermediate_size 16384 adopted) |
| Additional SFT | Korean-specialized Supervised Fine-Tuning |
| Developer | GinigenAI |
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)1@misc{ginigen_rogue_27b_kr_2026,
2 title = {Rogue-27B-KR: Korean-Specialized High-Performance Reasoning Model},
3 author = {GinigenAI},
4 year = {2026},
5 publisher = {Hugging Face},
6 howpublished = {\url{https://huggingface.co/ginigen-ai/Rogue-27B-KR}}
7}