kanana-2-vision-encoder-medium

🤗 kanana-2-vision-encoder-medium  |  🤗 kanana-2-vision-encoder-tiny  |  📕 Blog

kanana-2-vision-encoder-medium

News 🔥

• 2026/10/21: 🤗 Released the kanana-2-vision-encoder-medium and kanana-2-vision-encoder-tiny model weights.
• 2026/10/21: 📕 Published a blog post about the development of the kanana-2-vision-encoder series.

Introduction

Kanana-2-vision-encoder is Kakao's vision encoder series for vision-language models. This release includes medium (426M parameters) and tiny (98M parameters). The models extract image features for general visual understanding, document understanding, and visual reasoning, including tasks involving Korean text and culture. Both models are based on Kakao's Kanana-2-0.9B LLM and are feature-extraction-only models that exclude the text generation module. They are optimized as vision encoders for training vision-language models (VLMs) and perform best in this role.

Highlights

• Native Resolution: The encoder processes images at variable resolutions while preserving their aspect ratio.
• Efficient attention: Sliding Window Attention (SWA) and Grouped Query Attention (GQA) improve the efficiency of high-resolution image processing.

Model Downloads

Vision Encoder Performance

All comparisons use kanana-2-1.3b-instruct with a patch merger and the same training and evaluation conditions. Evaluation covers 42 benchmarks in nine categories. The average gives each category equal weight.

Medium (0.4B) Scale Comparison

Tiny (0.1B) Scale Comparison

† In-house benchmarks for Korean-language tasks.

§ Internally translated to Korean.

Deployment

pip install torch torchvision pillow transformers

import torch
from PIL import Image
from transformers import AutoImageProcessor, AutoModel

model_id = "kakaocorp/kanana-2-vision-encoder-medium"
device = "cuda" if torch.cuda.is_available() else "cpu"

processor = AutoImageProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModel.from_pretrained(model_id, trust_remote_code=True, dtype=torch.bfloat16)
model = model.to(device).eval()

images = [Image.open("example.jpg")]
inputs = processor(images, return_tensors="pt").to(device)
with torch.inference_mode():
    features = model(**inputs).last_hidden_state

print(features.shape)  # (total_patches, 768)

License

The model weights are released under the KananaOpenLicense.

Citation

@misc{kanana2visionencoder2026,
  title = {{kanana-2-vision-encoder}},
  author = {Jihwan Eom and Byungseok Roh},
  year = {2026},
  url = {https://huggingface.co/collections/kakaocorp/kanana-2-vision-encoder}
}

Contact

• Kanana MLLM Team Technical Support: kanana-mllm@kakaocorp.com