Kanana-2-v-decide explained: a 3B open decision model

By LLMTest Lab · Nov 6, 2026 · 4 min read decision-modelsopen-weightskakao
On this page

On this page

  1. What a decision model is
  2. What is new
  3. The claims, and who made them
  4. Who should care
  5. What we will test when we run it

Not tested by us yet. Kakao shipped weights only, with no hosted API we can call, so everything below comes from the model card on Hugging Face and Seoul Economic Daily's report. We'll update this post when we have run it on our own classification set.

Kanana-2-v-decide is a 3.67B-parameter model from Kakao that never writes a sentence. You give it a state (text, JSON, chat messages, optionally images) and typed questions, and it returns a probability for each possible answer. It is built for routing, classification and scoring steps inside agents, where an LLM call is overkill. It is open weight under the Kanana Open License, which Kakao says permits commercial use.

Kanana-2-v-decide-3b-2611
Maker Kakao Corp
Released 2026-11-06 (weights on Hugging Face)
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. No video
Question types choice, noul (yes/no), score (ordered scale)
Output Probabilities, plus a confidence value
License Kanana Open License
Price Free weights, you pay for your own GPU

What a decision model is

A chat model answers a question by generating tokens and you parse the result. A decision model skips generation. Per the model card, the state and each question are encoded into one embedding, every candidate answer into another, and a softmax over cosine similarities gives the answer distribution. confidence is the top probability minus the mean of the rest, and a temperature setting (default 1.0) flattens the distribution when raised.

That design has two practical consequences. First, the output is already structured, so there is no JSON that fails to parse. Second, the probabilities are relative to the candidates you supply: add or reword an option and the whole distribution shifts. The card says so directly, and it is the main thing to test before trusting a number. If you want background on why structured output matters in the first place, our guide to the best LLMs for structured output covers the generative alternative.

What is new

Kanana-2-v-decide sits on top of Kakao's own Kanana-2-v-3b, which pairs the Kanana-2-3b language model with the Kanana vision encoder. The news is the image input: the same request can judge a photo of a damaged product, the type of customer inquiry, and the customer's mood, then route the case. Kakao says it will first use the model internally, for web search and automatic product-category classification.

The card ships three ways to run it: transformers with trust_remote_code=True, a pipeline("decision", ...) interface, and a small HTTP server exposing POST /v1/systemone that is compatible with System One API clients. The server handles one request at a time and binds to 127.0.0.1 by default.

The claims, and who made them

Every number here is Kakao's, none is ours.

  • Speed: Seoul Economic Daily reports an average response time within 0.04 seconds on one Nvidia H200, about 43 times faster than GPT-6 Luna by Kakao's measurements. The model card itself makes no latency claim.
  • Accuracy: on the card, Kanana-2-v-decide scores 72.50 on Kakao's own Kanana Decision Bench (15,909 Korean text questions), against 54.38 for StartLux-Decision-v1.1-2B-Instant, 52.71 for Liquid AI d1-3B and 42.38 for Base Labs Joey-1. On the public Decision Index it scores 61.34, behind Joey-1 at 63.31, and it leads Decision Index Vision at 61.12. Its average of the two public sets is 61.23.

Two things to keep in mind. The 18-point lead is on a benchmark Kakao wrote, in Korean. On the public set the margin is small and Joey-1 wins the text score. Also, the news report lists a different comparison group (Microsoft's Decision-1, TypeSafe's Zeb, Perplexity's pplx-decider-v1.1-27B, Upstage's Solar Decide) than the model card does, so the two accounts are not directly reconcilable until someone reruns them.

Who should care

  • Agent builders paying for LLM calls on routing steps. A 3B local model with a fixed output shape can replace a prompt, a parser and a retry loop, if its accuracy holds on your labels.
  • Korean-language products. The headline benchmark is Korean, and Kakao's own use cases are Korean services.
  • Teams that cannot send data out. Weights run on your hardware, which matters for the privacy-sensitive cases in our list of open LLMs for on-prem deployments.

Skip it if you need explanations. It outputs a distribution, not reasoning. Each input also needs its own forward pass, overlong states are shortened (beginning and end kept), and EXIF image orientation is not applied, so normalize photos first.

What we will test when we run it

We plan to put it next to TypeSafe's Jev and a cheap and a mid-priced chat model acting as classifiers, on one shared set of 30 to 60 items. We will include control items with obvious answers and ask each yes/no question in both directions, because a decision model that flips its answer when the wording is mirrored is not usable for routing. Cost per 1,000 decisions and median latency go in the same table. Until then, treat the 0.04 second figure as a vendor claim measured on a data-center GPU, and the cost story as "free weights, your hardware". For the generative version of the same routing job, see cheap-first routing with escalation.

Related articles

Best LLMs for structured output in 2026: hit rates and retry costs
Which LLMs hit 99%+ JSON schema compliance natively in 2026, which need format enforcement middleware, and the real cost of retries at scale.
How to route LLM prompts in 2026: cheap first, escalate on fail
Route each prompt to the cheapest model that handles it well. When quality falls short, escalate silently. Here's the pattern with working Node.js code.