Instructions to use DFveloper/AIKAR-3-Pro-unquantized with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use DFveloper/AIKAR-3-Pro-unquantized with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("image-text-to-text", model="DFveloper/AIKAR-3-Pro-unquantized") 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, AutoModelForMultimodalLM processor = AutoProcessor.from_pretrained("DFveloper/AIKAR-3-Pro-unquantized") model = AutoModelForMultimodalLM.from_pretrained("DFveloper/AIKAR-3-Pro-unquantized") 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 Settings
- vLLM
How to use DFveloper/AIKAR-3-Pro-unquantized with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "DFveloper/AIKAR-3-Pro-unquantized" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "DFveloper/AIKAR-3-Pro-unquantized", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'Use Docker
docker model run hf.co/DFveloper/AIKAR-3-Pro-unquantized
- SGLang
How to use DFveloper/AIKAR-3-Pro-unquantized 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 "DFveloper/AIKAR-3-Pro-unquantized" \ --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": "DFveloper/AIKAR-3-Pro-unquantized", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'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 "DFveloper/AIKAR-3-Pro-unquantized" \ --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": "DFveloper/AIKAR-3-Pro-unquantized", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }' - Docker Model Runner
How to use DFveloper/AIKAR-3-Pro-unquantized with Docker Model Runner:
docker model run hf.co/DFveloper/AIKAR-3-Pro-unquantized
base_model: google/gemma-4-26b-a4b-it
tags:
- text-generation-inference
- transformers
- gemma4
license: apache-2.0
language:
- ko
๐ฐ๐ท AIKAR 3 Pro (26B) - Specialist in Korean Reasoning
AIKAR 3 Pro๋ LOOP์์ ๊ฐ๋ฐ๋ 26B ๊ท๋ชจ์ ์ธ์ด ๋ชจ๋ธ๋ก, ํนํ ํ๊ตญ์ด ์ถ๋ก (Reasoning) ๋ฅ๋ ฅ ๊ทน๋ํ๋ฅผ ์ํด ์ค๊ณ๋์์ต๋๋ค. ์ผ๋ฐ์ ์ธ ๋๊ท๋ชจ ์ธ์ด ๋ชจ๋ธ(LLM)์ด ๋ณด์ฌ์ฃผ๋ ๋จ์ ์ ๋ณด ์ ๊ณต์ ๋์ด, ๋ณต์กํ ํ๊ตญ์ด ๋ ผ๋ฆฌ ๊ตฌ์กฐ ์ดํด, ๋ฌธ๋งฅ์ ์ถ๋ก , ๋ค๋จ๊ณ ์ํ ๋ฐ ์ฝ๋ฉ ๋ฌธ์ ๋ฅผ ํ๊ตญ์ด ๋งฅ๋ฝ์์ ํ์ด๋ด๋ ๋ฐ ์ต์ ํ๋์ด ์์ต๋๋ค.
โจ Key Features (ํต์ฌ ๊ธฐ๋ฅ)
- Reasoning Focused (์ถ๋ก ์ค์ฌ): ๋จ์ ์์ฑ ๋ชจ๋ธ์ด ์๋, ๋ ผ๋ฆฌ์ ์ธ ๋จ๊ณ(Chain-of-Thought)๋ฅผ ๊ฑฐ์ณ ๋ต์ ๋์ถํ๋ ์ถ๋ก ํนํ ์ํคํ ์ฒ์ ๋๋ค. ํ๊ตญ์ด ๋ฌธ๋งฅ์์ ๋ฐ์ํ๋ ๋ฏธ๋ฌํ ๋์์ค๋ฅผ ๋ ผ๋ฆฌ ๊ตฌ์กฐ์ ๊ฒฐํฉํฉ๋๋ค.
- Korean-Centric Dataset (ํ๊ตญ์ด ํนํ): ํ๊ตญ์ด์ ๋ฌธ๋ฒ์ ํน์ฑ, ๋ฌธํ์ ๋ฐฐ๊ฒฝ, ์ ๋ฌธ ์ฉ์ด๋ฅผ ๊น์ด ์๊ฒ ์ดํดํ ์ ์๋๋ก ํ๊ตญ์ด ์ ์ ๋ฐ์ดํฐ์ ์ ์ค์ฌ์ผ๋ก ์ฌ์ ํ์ต(Pre-training) ๋ฐ ๋ฏธ์ธ ์กฐ์ (Fine-tuning)๋์์ต๋๋ค.
- Efficient 26B Architecture (26B ๊ท๋ชจ): ์ถ๋ก ๋ฅ๋ ฅ์ ํจ์จ์ฑ์ ๊ทน๋ํํ๊ธฐ ์ํด 26B ํ๋ผ๋ฏธํฐ๋ฅผ ์ฌ์ฉํ์ฌ, ์๋์ ์ผ๋ก ์ ์ VRAM์ผ๋ก๋ ๊ณ ์ฑ๋ฅ์ CoT(Reasoning) ๊ฒฐ๊ณผ๋ฅผ ์ป์ ์ ์๋๋ก ์ต์ ํ๋์์ต๋๋ค.
- Multi-task Capabilities: ์ํ์ ์ฌ๊ณ , ํ๋ก๊ทธ๋๋ฐ, ๋ฌธํ์ ์ถ๋ก , ๋ฒ๋ฅ ๋ฐ ๊ธฐ์ ๋ฌธ์ ํด์ ๋ฑ ๋ค์ํ ๊ณ ๋ํ๋ ์์ ์ ๋ฅ์ํฉ๋๋ค.
๐ ๋ชจ๋ธ ๊ตฌ์กฐ (Model Architecture)
AIKAR 3 Pro๋ ๋๊ท๋ชจ 26B ํ๋ผ๋ฏธํฐ ๋ ์ด์ด๋ฅผ ๊ฐ์ง ๋์ฝ๋ ์ ์ฉ ํธ๋์คํฌ๋จธ(Decoder-only Transformer) ๋ชจ๋ธ์ ๋๋ค. LOOP ๊ณ ์ ์ ์๊ณ ๋ฆฌ์ฆ์ ํตํด ํ๊ตญ์ด ํ ํฐ ์ฒ๋ฆฌ ํจ์จ์ 40% ์ด์ ํฅ์์์ผ, ๊ธด ๋ฌธ๋งฅ(Context window)์์๋ ์ถ๋ก ์ผ๊ด์ฑ์ ์ ์งํฉ๋๋ค.
- ํ๋ผ๋ฏธํฐ: 26 Billion
- Context Window: 32k tokens
- Focus: Korean Language Understanding, Logical Reasoning, Mathematical Solving
๐ ๏ธ ํ์ต ๊ณผ์ (Training Process)
AIKAR 3 Pro๋ ๋ค์ ์ธ ๊ฐ์ง ๋จ๊ณ๋ฅผ ๊ฑฐ์ณ ์์ฑ๋์์ต๋๋ค:
- Advanced Pre-training: ๋ฐฉ๋ํ ํ๊ตญ์ด ๋ฌธ๋ฒ ๊ต์ฌ, ๋ด์ค, ์ ๋ฌธ ์์ ๋ฐ ๊ณต๊ฐ ์น ๋ฐ์ดํฐ๋ฅผ ํ์ตํ์์ต๋๋ค.
- Supervised Fine-Tuning (SFT): ์ ๊ตํ๊ฒ ์ค๊ณ๋ ํ๊ตญ์ด ์ถ๋ก ํํ ๋ฆฌ์ผ ๋ฐ์ดํฐ์ ์ ํ์ตํ์ฌ ์๊ฐ์ ํ๋ฆ(Thought chain)์ ํ์ฑํ์ต๋๋ค.
- Reasoning Reinforcement Learning: ์ธ๊ฐ์ ์ ํธ๋๋ฅผ ๋ฐ์ํ ํ๊ตญ์ด ๋ ผ๋ฆฌ ๊ฒ์ฆ ๋ฃจํ๋ฅผ ํตํด, ๋จ์ํ ๋ต๋ง ๋ด๋ ๊ฒ์ด ์๋๋ผ ๋ ผ๋ฆฌ์ ์ผ๋ก ํ๋นํ ์ค๋ช (Rationales)์ ์ ๊ณตํ๋๋ก ์ต์ ํ๋์์ต๋๋ค.
๐ ์์ํ๊ธฐ (Getting Started)
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "DFveloper/AIKAR-3-Pro-unquantized"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype="auto")
prompt = """๋ฌธ์ ๊ฐ ์์ต๋๋ค. 37%์ ํ ์ธ์จ์ ์ ์ฉํ ์ํ์ด 15,000์์ผ ๋, ์๋ ๊ฐ๊ฒฉ์ ์ผ๋ง์
๋๊น? ๋จ๊ณ๋ณ๋ก ๋
ผ๋ฆฌ์ ์ผ๋ก ์ค๋ช
ํ์ธ์."""
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
๐ ๋ชจ๋ธ ์ฌ์ฉ ์ฃผ์์ฌํญ (Usage Note)
AIKAR 3 Pro๋ ์ถ๋ก ๋ฅ๋ ฅ์ ๊ฐํํ ๋ชจ๋ธ์ ๋๋ค. ๋ฐ๋ผ์ ๋ต์ ๋ฐ๋ก ์ป๊ธฐ๋ณด๋ค, "๋จ๊ณ์ ์ผ๋ก ์ค๋ช ํด์ค(Let's think step by step)" ๋๋ "๋ ผ๋ฆฌ์ ๊ณผ์ ์ ์์ธํ ์ ์ด์ค"์ ๊ฐ์ ํ๋กฌํํธ๋ฅผ ์ฌ์ฉํ ๋ ์ต์์ ์ฑ๋ฅ์ ๋ฐํํฉ๋๋ค.
๐ค ์ปค๋ฎค๋ํฐ ๋ฐ ์ฐ๋ฝ์ฒ
- Developers: LOOP Research Team
- Homepage: loop.ai (์์ ๋งํฌ)
- Report issues: [Github Issue Link]
ยฉ 2026 LOOP AIKAR Laboratory. All Rights Reserved.