Kanana

🤗 Kanana-2-v-decide-3b-2611  |  📕 Blog (Korean)

Kanana-2-v-decide-3b-2611

News 🔥

  • 2026/11/06: 🤗 Released the Kanana-2-v-decide-3b-2611 model weights.

Introduction

Kanana-2-v-decide-3b-2611 is Kakao's multimodal decision model for Korean and English. Given an input, it answers typed questions — selecting one of several options (choice), answering yes or no (noul), or rating on an ordered scale (score) — and returns a probability for each possible answer. It does not generate text. The input, called the state, may be text, a JSON object or array, or a chat history, optionally accompanied by images.

The model is built on Kanana-2-v-3b, which pairs the Kanana-2-3b language model with the Kanana vision encoder. Each question is encoded together with the state into a single embedding, and each candidate answer into another; a softmax over the cosine similarities between these embeddings yields the answer distribution.

Highlights

  • Public benchmarks: Achieves the highest average (61.23) across Decision Index and Decision Index Vision among the models compared, including the top score on Decision Index Vision.
  • Calibrated answers, no generation: Returns a full probability distribution over the supplied options for every question, with any number of questions per request.
  • Multimodal input: Accepts text, JSON, or chat messages with up to 10 images and up to 20,480 tokens.
  • Korean decisions: Scores 72.50 on Kanana Decision Bench (text), our Korean evaluation set — 18 points ahead of the next-best model compared.
Parameters 3.67B (vision encoder 0.49B, language model 3.18B)
Inputs Text, JSON, or chat messages; up to 10 images; up to 20,480 tokens
Questions choice, noul, score — any number per request
Embedding 2,560 dimensions

Model Downloads

Model Parameters Download
Kanana-2-v-decide-3b-2611 3.67B 🤗 HuggingFace

Decision Performance

Model Kanana Decision Bench† Public benchmarks Average
(public 2)
Text
15,909 questions
Decision Index
0.4.1 public
Decision Index Vision
0.4.1 public
Kanana-2-v-decide-3b-2611 (3B) 72.5061.3461.12 61.23
StartLux-Decision-v1.1-2B-Instant (2B) 54.3859.1759.90 59.54
Liquid AI d1-3B (3B) 52.7148.9751.86 50.42
Base Labs Joey-1 (2B) 42.3863.3154.23 58.77

The best score in each column is shown in bold. The average is calculated from the two displayed public benchmark scores.

† Kakao's in-house Korean decision evaluation set (text).

Deployment

Requirements

pip install -r requirements.txt

transformers

import torch
from transformers import AutoModel

model = AutoModel.from_pretrained("kakaocorp/kanana-2-v-decide-3b-2611", trust_remote_code=True,
                                  dtype=torch.bfloat16).to("cuda")

result = model.system_one(
    state="The item I ordered arrived damaged. Please replace it.",
    questions={
        "category": {"type": "choice", "instructions": "Classify the customer's inquiry.",
                     "criteria": {"delivery": "Delivery inquiry", "exchange": "Replacement request", "payment": "Payment inquiry"}},
        "damaged": {"type": "noul", "instructions": "Is the item damaged?"},
        "sentiment": {"type": "score", "instructions": "How positive is the customer's sentiment?",
                      "criteria": ["Negative", "Neutral", "Positive"]},
    },
)
print(result["answers"]["category"]["choice"])   # "exchange"
print(result["answers"]["damaged"]["noul"])      # probability of yes (0-1)
print(result["answers"]["sentiment"]["score"])   # expected level (0-2)

Images can be attached to the state as local paths, URLs, data:image/...;base64, URIs, or PIL images:

result = model.system_one(state="The customer attached a photo of the delivered item.",
                          images=["photo.jpg"],
                          questions={"damaged": {"type": "noul", "instructions": "Is the item damaged?"}})

The same request can also be made through a pipeline:

from transformers import pipeline

decide = pipeline("decision", model="kakaocorp/kanana-2-v-decide-3b-2611", trust_remote_code=True, device=0)
decide(state="The parcel never arrived.", questions={"refund": {"type": "noul", "instructions": "Does the customer want a refund?"}})

Serving over HTTP

server_kanana_v_decide.py exposes system_one through the System One API (POST /v1/systemone, GET /v1/models, GET /health) using only the Python standard library, so existing System One clients can connect simply by changing their base URL:

python server_kanana_v_decide.py --port 8000          # from the model directory, or --model <dir or Hub id>
curl -s localhost:8000/v1/systemone -H 'content-type: application/json' -d '{
  "state": "The parcel never arrived.",
  "questions": {"refund": {"type": "noul", "instructions": "Does the customer want a refund?"}}
}'

Requests accept the same inputs as system_one (state, questions, images, state_format, temperature, query_schemas); image inputs must be given as base64 data URLs. Requests are processed one at a time. To require Authorization: Bearer <key>, set DECISION_API_KEY. The server binds to 127.0.0.1 by default whether or not a key is set, so be sure to set a key before exposing it with --host.

Questions and answers

Type criteria Answer
choice A mapping from option keys to descriptions (a description may be null) choice (key), probabilities, confidence
noul Optional {"false": ..., "true": ...} descriptions noul: the probability that the statement is true
score A list of 2-10 ordered level descriptions, from low to high score: expected level index, probabilities, legend, confidence

confidence is the highest probability minus the mean of the remaining probabilities. Note that the probabilities are relative to the candidates supplied: adding or changing an option changes the distribution. temperature (default 1.0) divides the logits, so values above 1 produce a flatter distribution.

Additional system_one inputs:

  • state_format: one of "text", "json", or "messages". By default, strings are treated as text and objects or arrays as JSON; use "messages" for chat histories ([{"role": ..., "content": [...]}], containing text and image parts).
  • query_schemas: controls, for each question, which options are included in the query (null hides them).

Notes

  • By default, the options for choice and noul questions are included in the query; score levels are always included.
  • By default, overlong states are shortened while preserving their beginning and end; images are kept. Overlong candidate descriptions may also be shortened. Use --strict-inputs on the HTTP server to reject overlong inputs instead.
  • All images are converted to RGB: transparency in RGBA images is composited onto white, and EXIF orientation is not applied, so normalize the orientation of EXIF-tagged images beforehand. The processor first resizes images to at most 1,638,400 pixels, then enlarges small images to at least 65,536 pixels before splitting them into patches. Extreme aspect ratios are not supported. See processing_kanana_v_decide.py for details.
  • Each input is encoded in a separate forward pass. Candidate embeddings are cached across calls, keyed by token IDs.
  • Video input is not supported.

License

The model weights are released under the KananaOpenLicense.

Citation

@misc{kanana2vdecide2026,
  title = {{Kanana-2-v-decide}},
  author = {Sunghyo Chung and Juyoung Lee},
  year = {2026},
  url = {https://huggingface.co/kakaocorp/kanana-2-v-decide-3b-2611}
}

Contact

Downloads last month
1,362
Safetensors
Model size
4B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for kakaocorp/kanana-2-v-decide-3b-2611

Finetunes
1 model
Quantizations
1 model

Space using kakaocorp/kanana-2-v-decide-3b-2611 1

Collection including kakaocorp/kanana-2-v-decide-3b-2611