SSyncDev

Kanana-2 Vision Encoder

SyncDev model directory · Profile updated Nov 19, 2026

Compare with another model

Kakao's kanana-2-vision-encoder comes in 426M medium and 98M tiny sizes: open native-resolution image encoders for building Korean-strong VLMs.

Key specifications

Specification score
14.8
License
Kanana Open License

The specification score summarizes recorded features and limits. It is not a benchmark result.

Pricing conditions: Open weights for self-hosting; no hosted API price. Hardware cost is yours.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Image input

Produces

  • Embeddings output

Capabilities

VisionOpen weights

What is Kanana-2 Vision Encoder?

Kakao's kanana-2-vision-encoder series answers a narrow question that matters to anyone building a vision-language model: which image encoder should sit in front of the language model? Kakao released two answers on 21 October 2026, a 426M-parameter medium and a 98M-parameter tiny, both open and both tuned with Korean documents and culture in mind.

They extract features only. There is no text generation module; you attach your own LLM through a projector or patch merger.

Design choices

  • Native resolution: images keep their aspect ratio and are processed at variable size, which helps documents and receipts.
  • Sliding-window and grouped-query attention keep high-resolution inputs affordable.
  • Both encoders derive from Kakao's Kanana-2-0.9B language model.
  • Feature width per patch: 768 for medium, 512 for tiny.

How Kakao measured them

Every encoder was paired with Kanana-2 1.3B Instruct through a patch merger and trained and evaluated under identical conditions, across 42 benchmarks in nine categories, with each category weighted equally in the average.

Scale Encoder Average (9 categories)
~0.4B kanana-2-vision-encoder-medium (426M) 55.62
~0.4B PE-Lang-L14-448 (291M) 51.54
~0.4B Qwen3.5 encoder (415M) 50.51
~0.4B SigLIP2 so400m naflex (412M) 49.77
~0.1B kanana-2-vision-encoder-tiny (98M) 51.81
~0.1B Qwen3.5-0.8B encoder (88M) 49.62
~0.1B SigLIP2 base naflex (86M) 46.81

The biggest gains sit where you would expect: Korean OCR and documents (medium scores 86.80 on KoOCRBench), chart reading, and visual grounding, where medium reaches 79.88 on CountBenchQA. On some general English benchmarks such as MMBench the comparison encoders are level or ahead. These are Kakao's numbers.

One more detail stands out. Even the 98M tiny encoder beats every 0.4B rival in the table on average.

Using them

Install torch, torchvision, pillow and transformers, load with AutoImageProcessor and AutoModel with custom code, and you get a tensor of shape (total patches, feature dim). Both repos carry the same documentation, so the choice is purely about size.

Picking a size

Tiny suits on-device or high-throughput VLMs; medium suits document-heavy assistants. Run both behind the same small LLM on a sample of your real images, especially Korean receipts or forms if that is your market, and compare OCR accuracy before committing to a full VLM training run.

Weights are under the Kanana Open License. Kakao's Kanana-2 V Decide 3B shows the family's multimodal direction; more encoders are in the image feature extraction directory.

How to evaluate Kanana-2 Vision Encoder

Compare the model on the work you need to complete.

Tasks and inputs

Kanana-2 Vision Encoder is listed for Image feature extraction, Document OCR. It accepts image and produces embedding. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Evaluate retrieval over a labelled query set. Keep the encoder, vector dimensions and index settings consistent.

Choosing a deployment

The weights are downloadable under Kanana Open License. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

Medium with 426M parameters and tiny with 98M parameters, producing 768- and 512-dimensional patch features respectively.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

allenai

OlmoEarth v1.3 Base

OlmoEarth v1.3 Base is Ai2's open Earth-observation model that turns a year of Sentinel and Landsat imagery into a 128-dimension embedding per 10 m pixel.

— context

camenduru

camenduru/dinov3-vitl16-pretrain-lvd1689m

camenduru/dinov3-vitl16-pretrain-lvd1689m is a source-linked Hugging Face repository assigned to Image feature extraction. Its recorded task interface accepts image and produces embeddings.

— context

facebook

facebook/dinov2-base

facebook/dinov2-base is a source-linked Hugging Face repository assigned to Image feature extraction. Its recorded task interface accepts image and produces embeddings.

— context