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

kanana-2-vision-encoder-tiny

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

Model	Download
kanana-2-vision-encoder-medium	🤗 HuggingFace
kanana-2-vision-encoder-tiny	🤗 HuggingFace

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

Benchmark	apple/ aimv2-large- patch14-native (308M)	facebook/ PE-Lang- L14-448 (291M)	google/ siglip2-so400m- patch16-naflex (412M)	Qwen/ Qwen3.5-397B- A17B (415M)	kakaocorp/ kanana-2-vision- encoder-medium (426M)
STEM Puzzle
MMMU (Val)	37.56	38.85	38.52	38.15	39.70
MMMU-Pro	22.27	22.99	23.14	22.62	22.66
MathVista (testmini)	44.27	47.47	45.73	46.37	51.30
MathVerse (testmini)	24.22	24.91	24.34	24.61	27.49
MathVision	14.77	14.51	14.36	15.77	15.32
DynaMath	26.12	28.42	26.15	28.06	30.37
WeMath	12.03	12.00	12.32	11.43	14.48
LogicVista	28.26	31.77	27.67	29.68	29.16
VisualPuzzles	27.11	25.40	25.77	26.60	26.26
ScienceQA (Val)	72.07	74.38	71.79	71.53	75.04
General VQA
MMBench (EN)	59.94	60.62	61.00	58.02	58.42
RealWorldQA	49.89	56.38	55.25	55.08	59.22
MMStar	43.71	45.91	47.62	45.56	49.33
WildVision	38.87	44.80	41.87	39.53	40.73
Alignment
HallusionBench	36.73	37.17	38.27	39.76	38.83
MIA-Bench	81.04	80.25	76.52	80.44	77.10
MIA-Bench (Korean)§	81.06	82.85	83.14	81.55	81.75
MM-IFEval	37.12	37.02	36.80	38.72	37.84
Document Understanding
OCRBench	60.90	69.60	62.50	68.80	75.60
CharXiv (DQ)	49.24	57.08	53.09	54.84	61.19
CharXiv (RQ)	19.13	24.20	19.77	22.30	28.63
Fine-grained Perception
V*	44.85	55.67	55.67	58.12	56.37
HR-Bench (4K)	45.79	55.75	53.50	52.04	55.38
HR-Bench (8K)	41.00	46.83	44.92	45.50	46.71
Visual Grounding
CountBenchQA	55.17	57.77	61.05	54.62	79.88
RefCOCO (Avg.)	48.73	61.57	61.12	59.73	76.14
Korean Exam
KoNET	44.16	46.78	43.49	44.03	50.58
KoMathSolution†	30.92	33.29	32.38	31.42	33.79
KoExamGrad†	59.82	45.86	32.97	54.14	67.86
KoExamCSAT†	22.56	21.29	21.24	19.33	22.12
Korean OCR/Doc. Understanding
KoPDFBench†	48.22	56.14	52.62	56.05	56.77
KoOCRBench†	63.41	79.42	69.81	75.89	86.80
KoReceiptQA†	51.24	54.73	50.24	52.25	62.83
KoChart†	65.21	71.99	69.39	69.33	79.07
KoCosMed†	67.98	77.29	74.38	75.25	78.75
KoFoodMenu†	53.97	60.80	60.39	60.87	65.15
Korean Culture
KoEntity†	54.95	55.79	55.54	56.64	59.23
KoCelebQA†	32.39	33.46	32.30	33.03	32.60
KoCelebRecognition†	42.92	46.15	43.02	45.73	45.62
KoMoviePosterQA†	52.71	62.44	59.28	60.21	67.16
KoFoodQA†	66.56	69.93	75.42	69.79	70.41
K-Viscuit	57.69	57.84	57.79	55.76	57.84
Average across 9 categories	47.31	51.54	49.77	50.51	55.62

Tiny (0.1B) Scale Comparison

Benchmark	apple/ MobileCLIP2-B (86M)	facebook/ PE-Core- B16-224 (86M)	google/ siglip2-base- patch16-naflex (86M)	Qwen/ Qwen3.5-0.8B (88M)	kakaocorp/ kanana-2-vision- encoder-tiny (98M)
STEM Puzzle
MMMU (Val)	37.70	37.74	38.37	39.67	39.00
MMMU-Pro	19.31	20.44	21.50	20.67	22.14
MathVista (testmini)	44.00	45.43	44.20	48.17	49.00
MathVerse (testmini)	22.45	24.33	23.83	25.74	25.51
MathVision	14.73	15.34	15.55	15.21	14.85
DynaMath	23.31	25.05	25.09	29.28	29.27
WeMath	9.94	11.59	10.98	13.55	13.14
LogicVista	27.14	29.83	28.04	29.68	28.41
VisualPuzzles	26.60	24.86	25.49	26.86	25.91
ScienceQA (Val)	70.85	70.78	71.94	72.75	72.23
General VQA
MMBench (EN)	55.24	55.44	54.50	56.10	56.98
RealWorldQA	51.15	50.20	56.30	55.73	57.82
MMStar	43.31	42.91	44.04	45.58	45.44
WildVision	30.73	34.80	34.40	38.07	37.07
Alignment
HallusionBench	35.37	36.58	38.02	39.77	36.36
MIA-Bench	79.40	79.35	79.02	76.83	81.01
MIA-Bench (Korean)§	76.75	82.21	75.72	77.80	80.92
MM-IFEval	38.05	37.53	37.62	38.19	37.76
Document Understanding
OCRBench	44.63	59.30	52.53	67.57	66.70
CharXiv (DQ)	43.91	50.88	50.00	56.47	57.52
CharXiv (RQ)	15.37	20.73	19.30	24.40	25.03
Fine-grained Perception
V*	48.69	52.36	49.39	55.15	55.36
HR-Bench (4K)	47.75	48.17	50.50	52.79	53.75
HR-Bench (8K)	41.38	41.83	44.75	46.92	45.92
Visual Grounding
CountBenchQA	49.49	46.48	50.51	53.25	69.47
RefCOCO (Avg.)	53.74	57.13	54.85	58.95	66.07
Korean Exam
KoNET	31.30	42.00	39.94	45.72	46.29
KoMathSolution†	29.92	27.50	30.46	30.04	31.21
KoExamGrad†	27.09	30.95	49.62	38.18	58.65
KoExamCSAT†	19.82	20.22	20.12	21.10	19.73
Korean OCR/Doc. Understanding
KoPDFBench†	44.87	51.93	47.90	53.87	51.64
KoOCRBench†	62.21	69.05	59.52	79.34	80.57
KoReceiptQA†	49.59	50.30	48.23	52.90	53.07
KoChart†	59.95	68.48	64.31	65.40	73.02
KoCosMed†	67.83	72.85	70.60	75.76	74.24
KoFoodMenu†	54.87	56.31	57.35	58.25	61.08
Korean Culture
KoEntity†	51.73	50.40	52.07	53.43	52.35
KoCelebQA†	29.30	32.43	30.54	31.62	31.62
KoCelebRecognition†	44.02	44.43	43.57	44.70	44.42
KoMoviePosterQA†	55.50	55.08	58.44	63.04	62.79
KoFoodQA†	66.59	61.98	67.79	65.69	66.98
K-Viscuit	52.31	52.97	55.76	54.90	55.21
Average across 9 categories	44.20	46.60	46.81	49.62	51.81

† 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-tiny"
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, 512)

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