Text Generation
Transformers
Safetensors
Korean
English
qwen3_5
image-text-to-text
korean
darwin
darwin-platform
Merge
conversational
Instructions to use Anserwise/AWAXIS-Hybrid-28B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use Anserwise/AWAXIS-Hybrid-28B with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="Anserwise/AWAXIS-Hybrid-28B") messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"}, {"type": "text", "text": "What animal is on the candy?"} ] }, ] pipe(text=messages)# Load model directly from transformers import AutoProcessor, AutoModelForImageTextToText processor = AutoProcessor.from_pretrained("Anserwise/AWAXIS-Hybrid-28B") model = AutoModelForImageTextToText.from_pretrained("Anserwise/AWAXIS-Hybrid-28B") messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"}, {"type": "text", "text": "What animal is on the candy?"} ] }, ] inputs = processor.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt", ).to(model.device) outputs = model.generate(**inputs, max_new_tokens=40) print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:])) - Notebooks
- Google Colab
- Kaggle
- Local Apps
- vLLM
How to use Anserwise/AWAXIS-Hybrid-28B with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "Anserwise/AWAXIS-Hybrid-28B" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Anserwise/AWAXIS-Hybrid-28B", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/Anserwise/AWAXIS-Hybrid-28B
- SGLang
How to use Anserwise/AWAXIS-Hybrid-28B with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "Anserwise/AWAXIS-Hybrid-28B" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Anserwise/AWAXIS-Hybrid-28B", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "Anserwise/AWAXIS-Hybrid-28B" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Anserwise/AWAXIS-Hybrid-28B", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }' - Docker Model Runner
How to use Anserwise/AWAXIS-Hybrid-28B with Docker Model Runner:
docker model run hf.co/Anserwise/AWAXIS-Hybrid-28B
Update README: humble tone, remove competitive claims
Browse files
README.md
CHANGED
|
@@ -10,61 +10,55 @@ base_model:
|
|
| 10 |
- FINAL-Bench/Darwin-28B-KR
|
| 11 |
tags:
|
| 12 |
- korean
|
| 13 |
-
- awaxis
|
| 14 |
- darwin
|
| 15 |
- darwin-platform
|
| 16 |
-
- reasoning
|
| 17 |
-
- hybrid
|
| 18 |
- merge
|
| 19 |
---
|
| 20 |
|
| 21 |
# AWAXIS-Hybrid-28B
|
| 22 |
|
| 23 |
-
>
|
| 24 |
|
| 25 |
-
AWAXIS-Think-28B
|
| 26 |
|
| 27 |
-
|
| 28 |
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
- **🧬 아버지**: [`Anserwise/AWAXIS-Think-28B`](https://huggingface.co/Anserwise/AWAXIS-Think-28B) — Reasoning 강점
|
| 32 |
-
- **🧬 어머니**: [`FINAL-Bench/Darwin-28B-KR`](https://huggingface.co/FINAL-Bench/Darwin-28B-KR) — 한국어 도메인
|
| 33 |
-
|
| 34 |
-
## ✨ Highlight
|
| 35 |
-
|
| 36 |
-
K-AI 환산 **0.560** (Quetta-V3 1위 동률 예상) · KMMLU-Pro 0.674 · MuSR(Ko) 0.611
|
| 37 |
|
| 38 |
---
|
| 39 |
|
| 40 |
## 📊 평가 결과
|
| 41 |
|
| 42 |
-
### 1) K-AI 리더보드 기준 (5과목
|
| 43 |
|
| 44 |
KMMLU-Pro / CLIcK / HLE(Ko) / MuSR(Ko) / Com2-main(ko)
|
| 45 |
-
- **Quetta-V3, Rogue-28B-MIX**: K-AI 리더보드 실제 측정값 (leaderboard.aihub.or.kr)
|
| 46 |
-
- **AWAXIS-Hybrid-28B**: 자체 mirror eval(100문항) × Rogue-MIX 기준 ratio 환산 예상값
|
| 47 |
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
| 🥇 동률 | **AWAXIS-Hybrid-28B** ⭐ (이 모델, 예상) | 0.674 | 0.787 | 0.07 | 0.611 | 0.657 | **2.799** | **0.560** |
|
| 51 |
-
| 🥇 동률 | Quetta-V3 (실측) | 0.676 | 0.794 | 0.07 | 0.604 | 0.654 | **2.798** | **0.560** |
|
| 52 |
-
| 🥉 3 | Rogue-28B-MIX (실측) | 0.666 | 0.797 | 0.07 | 0.611 | 0.650 | **2.794** | **0.559** |
|
| 53 |
-
| 4 | Warecube-KO-27B-v3 (실측) | 0.668 | 0.799 | 0.067 | 0.584 | 0.638 | 2.756 | 0.551 |
|
| 54 |
-
| 5 | AWAXIS-Think-28B (실측) | 0.603 | 0.770 | 0.055 | 0.591 | 0.632 | 2.651 | 0.530 |
|
| 55 |
|
| 56 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 57 |
|
| 58 |
-
|
|
|
|
|
|
|
|
|
|
| 59 |
|
| 60 |
CLIcK + KMMLU(history/law/health) + HAERAE(gk/hist/ling) + KOBEST(hella/copa/boolq)
|
| 61 |
|
| 62 |
-
| Model | CLIcK | KMMLU 평균 | HAERAE 평균 | KOBEST 평균 | **Sum (10)** | **Macro** |
|
| 63 |
|-------|:---:|:---:|:---:|:---:|:---:|:---:|
|
| 64 |
-
| **AWAXIS-Hybrid-28B** ⭐ | 0.83 | 0.530 | 0.813 | 0.967 | **7.760** | **0.7760** |
|
| 65 |
-
|
|
| 66 |
|
| 67 |
-
> KMMLU 평균에서 본 모델이 +0.033 (도메인 지식 우세), HAERAE에서 -0.024 (한국어 표면 패턴)
|
| 68 |
|
| 69 |
---
|
| 70 |
|
|
@@ -87,34 +81,9 @@ outputs = model.generate(inputs, max_new_tokens=256, temperature=0.0)
|
|
| 87 |
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))
|
| 88 |
```
|
| 89 |
|
| 90 |
-
## 머지 레시피
|
| 91 |
-
|
| 92 |
-
Smart MRI Layer-wise — Layer 그룹별로 부모 가중치 비율을 다르게 적용
|
| 93 |
-
|
| 94 |
-
| 영역 | 어머니(Darwin-28B-KR) 채택률 | 의도 |
|
| 95 |
-
|------|:---:|------|
|
| 96 |
-
| Embed / LM-head | 50% | 출력 통로 균형 |
|
| 97 |
-
| Norm | 30% | 안정성 위주 |
|
| 98 |
-
| Visual encoder | 0% | 아버지 보존 |
|
| 99 |
-
| Layers 0~15 (초기) | 40% | 한국어 표면 패턴 |
|
| 100 |
-
| Layers 16~50 (중기) | 0% | Reasoning 능력 보존 |
|
| 101 |
-
| Layers 51~63 (후기) | 70% | 도메인 지식 채택 |
|
| 102 |
-
|
| 103 |
## 라이선스
|
| 104 |
|
| 105 |
Apache 2.0 (부모 모델 라이선스 계승)
|
| 106 |
|
| 107 |
-
## 인용
|
| 108 |
-
|
| 109 |
-
```bibtex
|
| 110 |
-
@misc{awaxis_hybrid_28b_2026,
|
| 111 |
-
title={AWAXIS-Hybrid-28B: AWAXIS-Think × Darwin Platform Korean LLM},
|
| 112 |
-
author={Anserwise · ginigen-ai · VIDraft team},
|
| 113 |
-
year={2026},
|
| 114 |
-
publisher={Hugging Face},
|
| 115 |
-
url={https://huggingface.co/Anserwise/AWAXIS-Hybrid-28B}
|
| 116 |
-
}
|
| 117 |
-
```
|
| 118 |
-
|
| 119 |
---
|
| 120 |
-
*
|
|
|
|
| 10 |
- FINAL-Bench/Darwin-28B-KR
|
| 11 |
tags:
|
| 12 |
- korean
|
|
|
|
| 13 |
- darwin
|
| 14 |
- darwin-platform
|
|
|
|
|
|
|
| 15 |
- merge
|
| 16 |
---
|
| 17 |
|
| 18 |
# AWAXIS-Hybrid-28B
|
| 19 |
|
| 20 |
+
> AWAXIS-Think × Darwin Platform Hybrid 모델
|
| 21 |
|
| 22 |
+
AWAXIS-Think-28B를 기반으로 Darwin Platform 한국어 가중치를 Smart MRI Layer-wise 머지로 결합한 한국어 LLM입니다.
|
| 23 |
|
| 24 |
+
본 모델은 **Anserwise**에서 제작·공개한 한국어 LLM입니다.
|
| 25 |
|
| 26 |
+
- **🧬 아버지**: [`Anserwise/AWAXIS-Think-28B`](https://huggingface.co/Anserwise/AWAXIS-Think-28B)
|
| 27 |
+
- **🧬 어머니**: [`FINAL-Bench/Darwin-28B-KR`](https://huggingface.co/FINAL-Bench/Darwin-28B-KR)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 28 |
|
| 29 |
---
|
| 30 |
|
| 31 |
## 📊 평가 결과
|
| 32 |
|
| 33 |
+
### 1) K-AI 리더보드 기준 (5과목)
|
| 34 |
|
| 35 |
KMMLU-Pro / CLIcK / HLE(Ko) / MuSR(Ko) / Com2-main(ko)
|
|
|
|
|
|
|
| 36 |
|
| 37 |
+
- **외부 모델**: K-AI 리더보드(leaderboard.aihub.or.kr) 실측값
|
| 38 |
+
- **본 시리즈**: 자체 mirror eval(100문항) × Rogue-28B-MIX 기준 ratio 환산 추정값
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 39 |
|
| 40 |
+
| Model | KMMLU-Pro | CLIcK | HLE(Ko) | MuSR(Ko) | Com2-main(ko) | **Sum** | **Macro** |
|
| 41 |
+
|-------|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
|
| 42 |
+
| Hybrid (예상) | 0.674 | 0.787 | 0.07 | 0.611 | 0.657 | **2.799** | **0.560** |
|
| 43 |
+
| **AWAXIS-Hybrid-28B** ⭐ (이 모델, 예상) | 0.674 | 0.787 | 0.07 | 0.611 | 0.657 | **2.799** | **0.560** |
|
| 44 |
+
| Rogue-28B-MIX (실측) | 0.666 | 0.797 | 0.07 | 0.611 | 0.650 | **2.794** | **0.559** |
|
| 45 |
+
| Warecube-KO-27B-v3 (실측) | 0.668 | 0.799 | 0.07 | 0.584 | 0.638 | **2.756** | **0.551** |
|
| 46 |
+
| AWAXIS-Think-28B (실측) | 0.603 | 0.770 | 0.06 | 0.591 | 0.632 | **2.651** | **0.530** |
|
| 47 |
+
| KR-Pro (예상) | 0.643 | 0.661 | 0.07 | 0.585 | 0.650 | **2.609** | **0.522** |
|
| 48 |
+
| KR-Plus (예상) | 0.643 | 0.703 | 0.07 | 0.532 | 0.657 | **2.605** | **0.521** |
|
| 49 |
|
| 50 |
+
|
| 51 |
+
> HLE(Ko)는 28B급 공통 약점 (난이도 매우 높음).
|
| 52 |
+
|
| 53 |
+
### 2) 종합 한국어 능력 (10과목 mirror eval)
|
| 54 |
|
| 55 |
CLIcK + KMMLU(history/law/health) + HAERAE(gk/hist/ling) + KOBEST(hella/copa/boolq)
|
| 56 |
|
| 57 |
+
| Model | CLIcK | KMMLU 평균 | HAERAE 평균 | KOBEST 평균 | **Sum (10과목)** | **Macro** |
|
| 58 |
|-------|:---:|:---:|:---:|:---:|:---:|:---:|
|
| 59 |
+
| **AWAXIS-Hybrid-28B** ⭐ (이 모델) | 0.83 | 0.530 | 0.813 | 0.967 | **7.760** | **0.7760** |
|
| 60 |
+
| Rogue-28B-MIX | 0.83 | 0.513 | 0.807 | 0.967 | **7.690** | **0.7690** |
|
| 61 |
|
|
|
|
| 62 |
|
| 63 |
---
|
| 64 |
|
|
|
|
| 81 |
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))
|
| 82 |
```
|
| 83 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 84 |
## 라이선스
|
| 85 |
|
| 86 |
Apache 2.0 (부모 모델 라이선스 계승)
|
| 87 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 88 |
---
|
| 89 |
+
*2026-04-30*
|