Kakao Kanana-2-v-decide: a 3.67B decision model that answers in probabilities, not prose

Kakao has released a model that refuses to write. Kanana-2-v-decide, published on Hugging Face on November 6, takes text, JSON or chat — plus up to 10 images — and answers typed questions with probabilities, not prose: pick an option, answer yes or no, or rate on a scale. Kakao positions it as the decision layer for AI agents, the small component that classifies a request and picks the next tool.

The model, published on Hugging Face as Kanana-2-v-decide-3b-2611, is built on Kanana-2-v-3b, Kakao’s in-house multimodal language model, and carries 3.67 billion parameters: a 0.49-billion-parameter vision encoder paired with a 3.18-billion-parameter language model. It accepts text, JSON or chat history as its input state, with up to 10 images and up to 20,480 tokens per request.

How it decides

Instead of generating a reply token by token, the model encodes each question together with the input state into a single embedding, and each candidate answer into another. A softmax over the cosine similarities between those embeddings produces the answer distribution — a probability for every option, every time. Candidate embeddings are cached across calls, so repeated choices get cheaper.

There are three question types. Choice picks one of several supplied options. Noul answers yes or no, returning the probability that a statement is true. Score rates on an ordered scale of two to ten levels. The same request can ask any number of questions at once, and the answers come back structured, so downstream systems never have to parse free-form prose to find the conclusion.

Kakao’s worked example is customer support: feed the model a message saying an item arrived damaged, plus a photo of it, and ask whether it is damaged and what the inquiry is about. It returns a probability the item is damaged and a category — exchange, delivery, payment — that a service can weigh against its own thresholds to approve a replacement or route the case to a human.

Fast, and small on purpose

The headline number is speed. Kakao says the model averages under 0.04 seconds per response on a single Nvidia H200 GPU. That is the point of the design: a decision model is a component, not a chatbot, and agents that call it dozens of times per task need the answer in milliseconds, not seconds.

On benchmarks, Kakao reports 72.50 on its own Kanana Decision Bench, a set of 15,909 Korean text questions, which it says is 18 points ahead of the next-best model it compared. On the public Decision Index benchmarks it scores 61.34 on text and 61.12 on vision, for a 61.23 average — the highest average among the four models in Kakao’s comparison table, which also includes StartLux Decision, Liquid AI’s d1-3B and Base Labs’ Joey-1. The caveat is worth stating plainly: Joey-1 leads on the text-only Decision Index (63.31 against 61.34), and all of these figures are Kakao’s own reporting.

How to use it

The weights are on Hugging Face under the Kanana Open License, which permits commercial use. Developers can load the model through the Transformers pipeline for decision tasks, call its system_one method directly, or run the small HTTP server Kakao ships with the repo, which exposes the same interface over a local endpoint. Kakao also published a Korean-language technical post on the release.

Kakao’s first use is internal: the company says it will apply the model to its own web search and automatic product categorization, and later fold search and decision-making into a single model. The release lands a day after Kakao’s third-quarter earnings call, where chief executive Chung Shin-a said the company will begin piloting Kanana Space, an AI execution space inside KakaoTalk, in December, and formally launch two PC-based services, Kanvas and Modu-ui AI, before the end of the year — part of a push toward 10 million monthly AI users by year-end. Small, fast decision models are the plumbing for that plan: every agent action starts with a choice, and Kakao now sells the component that makes it.

The decision-model field is already crowded — Microsoft, Upstage and a clutch of startups ship similar components — but Kakao’s entry is notable for what it refuses to do. Where most small models chase chat, this one only decides. For agent builders, that narrowness is the feature: a 3.67B model that returns a probability distribution in 40 milliseconds is easier to trust, and cheaper to run, than a general model asked to reason its way to the same answer.