[Edaily Reporter Lee So-Hyun ] Kakao(035720)has independently developed a vision encoder—which serves as the “eyes” of multimodal artificial intelligence (AI)—and released it as open source. The company stated that even its lightweight models, with fewer than 100 million parameters, demonstrated high performance in evaluations comparing them to leading global vision encoders from companies such as Google, Meta, Apple, and Alibaba.

Kakao announced on the 21st that it has released two versions of its self-developed “Kanana-2-vision-encoder” on the global AI model platform Hugging Face.

The released models are a “medium” model with 426 million parameters and a “tiny” model with 98 million parameters. Kakao had previously announced plans to release the vision encoders at its developer conference, “IfKakao26,” held on the 13th.

The AI’s “Eyes” for Understanding Images and Documents

The vision encoder extracts visual information—such as objects, text, location, and shape—from images and passes it to the language model.

The language model interprets this information alongside the user’s question to generate a response. It can be used to read food items and prices on a menu, understand the content and layout of a document, and describe the scene in a photo.

Kakao evaluated the performance of the two models across a total of 42 benchmarks in nine areas, including general visual question-answering (VQA), English and Korean text recognition and document understanding, and image localization and detailed recognition.

The company explained that the evaluation results showed the models achieved higher overall performance than major global vision encoders, such as Google’s “SigLIP 2,” Meta’s “Perception Encoder,” Apple’s “AIMv2,” and Alibaba’s “Qwen 3.5-ViT.”

In particular, Kakao explained that the “tiny” model, despite having fewer than 100 million parameters, demonstrated competitive performance even when compared to global models with over 400 million parameters.

The company noted that the performance gap was particularly pronounced in the areas of image localization, Korean language testing, Korean character recognition, and document understanding.

Kakao explained that it enhanced its optical character recognition (OCR) and document understanding performance by applying its proprietary learning technology, which preserves visual information from small text and complex document structures.

4x Training Efficiency… Over 60% Reduction in Processing Latency

Kakao announced that it increased the training efficiency of its vision encoder by approximately four times compared to previous levels by utilizing its in-house language model, “Kanana-2.”

The company explained that it leveraged the language model’s capabilities in training the vision encoder to extract detailed visual information from high-resolution images.

By applying a design optimized for high-resolution image processing, the company also reduced processing latency by more than 60% compared to leading global vision encoders.

Additionally, by applying knowledge distillation and lightweighting techniques, the company developed not only a medium-sized model but also a lightweight model designed for on-device use.

Commercial Use Permitted Beyond Research

Kakao is applying the “Kanana Open License” to both models.

The plan is to allow not only research and development but also commercial use, enabling domestic and international researchers and companies to utilize the technology for the development of multimodal AI technologies and services.

Kakao is independently developing the components that make up the omnichannel language model “Kanana-o,” including the vision encoder, language models, and visual and speech processing modules. The company plans to release other components sequentially in the future.

Noh Byung-seok, Performance Leader for Kakao’s Unified Foundation Model, stated, “In the era of agentic AI, ‘computer use’ technology—which directly understands and manipulates PC or mobile screens—is gaining attention, and the role of vision encoders that accurately analyze high-resolution images is growing.” He added, “We will enhance the competitiveness of our proprietary AI models and contribute to the development of the AI ecosystem by releasing related technologies.”