The News, Explained
Kakao released medium and tiny versions of Kanana-2-vision-encoder on October 21, 2026. Official descriptions list 426 million and 98 million parameters, respectively. Parameters are numerical values learned within a model; these models extract image features. Source Source Source
Reading an image and generating an answer
A vision encoder turns visual information, such as objects, text, positions and shapes, into numerical representations a language model can use. The cards identify these as feature-extraction-only models without a text generation module, optimized for training vision-language models. A document-question-answering system needs to connect the extracted features to a language model and other components that interpret them with the question. Source Source Source
The cards describe variable-resolution processing that preserves image aspect ratio, a relevant property when assessing small text and layouts. Kakao presents tiny as a lightweight model developed with on-device use in mind. Memory use and processing speed on an intended device remain items to evaluate separately. Source Source Source
The conditions behind performance comparisons
Kakao’s comparison connects each encoder to the same kanana-2-1.3b-instruct language model and a patch merger that combines image features, under the same training and evaluation conditions. It covers 42 benchmarks across nine categories, with each category weighted equally in the average. Some Korean benchmarks were made in-house. Source Source
Results differ by task. In the medium-scale table, Kanana medium scores 75.60 on OCRBench, a text-recognition and document-understanding test, compared with 62.50 for SigLIP 2. On MMBench English visual question answering, the respective scores are 58.42 and 61.00. Even within the matched setup, the leading model varies with the question. These are Kakao’s published results, not independent OYOPICK measurements. Source
Commercial use and supplying models
Both models use the Kanana Open License. Section 4.2 allows development and operation of one’s own services without a separate commercial license outside the activities in section 4.1, subject to the agreement’s other conditions. Source Source
Section 4.1 requires a separate commercial license from Kakao for offering or reselling third-party access to the model or derivatives through APIs, cloud platforms or remote services; supplying them through system-integration or on-premise solutions; or offering or reselling them embedded in devices. These terms distinguish using a model in one’s service from supplying the model or access to others. Source Source
Redistribution conditions include passing on use restrictions, providing the agreement, marking modifications, including a Notice file and displaying “Powered by Kanana.” A Kanana naming prefix also applies when distributing AI models created, trained or improved using the materials or outputs. The published agreement provides the full scope. Source Source
OYOPICK’s Take
The useful addition is another component to assess separately in AI that handles Korean documents and images. When document question answering goes wrong, a team may need to distinguish misread small text from a mistake in interpreting correctly extracted information. Access to the visual component could help that investigation.
The strongest selection criterion is the actual documents and questions a system will handle. The published table already shows task-dependent rankings, while supplying a model in different ways changes the license conditions. Evaluating recognition and answers together on relevant material, alongside the intended service’s conditions, could turn this release into a more concrete product choice.