A model with nothing to say
The design is the story. A conventional language model asked to classify a customer message produces a sentence, and the calling system then has to parse that sentence back into a category. Kanana-2-v-decide skips both steps. You supply the state and a set of questions with defined answers. It encodes the question together with the state, encodes each candidate answer separately, and returns a probability distribution from the similarity between the two. The card's own example is a damaged-delivery message that comes back as a category (exchange), a damage probability, and a sentiment score in one call.
That makes it infrastructure rather than a standalone conversational product. Its intended role is to handle the small, constant judgments inside agent systems: which tool to call next, whether a condition holds, and how severe a case is. Kakao says it will use the model first on its own web search and on classifying product categories, which is the right scale of ambition for a 3.67B model designed to return structured judgments rather than natural-language explanations.
Two limits sit plainly on the card. Video input is not supported, which rules out a large class of multimodal checking. And the reference server Kakao ships with the model processes requests one at a time. The reference HTTP server processes requests sequentially. Teams deploying it at scale will need to assess their own serving architecture, concurrency requirements and throughput.
Reading the table Kakao published
Kakao's headline is that the model achieves the highest average across the public Decision Index and Decision Index Vision sets among the four models included in its comparison. That is true of the table, and the table is also where the claim gets smaller.
The average lead rests on the vision set. On Decision Index Vision, Kanana scores 61.12, ahead of StartLux at 59.90. On the text-only Decision Index, Base Labs' Joey-1, a 2B model, scores 63.31 against Kanana's 61.34. Averaging the two sets puts Kanana first at 61.23, Joey-1 second at 58.77. Both statements are accurate. Only one of them fits in a headline.
The comparison set is also narrow: four models, all small, all purpose-built deciders (StartLux-Decision-v1.1-2B-Instant, Liquid AI's d1-3B, Base Labs' Joey-1). This is a ranking inside a young niche, not a result against general-purpose models running the same questions.
The Korean figure needs the largest discount. Kanana Decision Bench is Kakao's own evaluation set, 15,909 text questions, and on it Kanana scores 72.50, more than 18 points ahead of the next model. A model developed by Kakao scoring strongly on a Korean-language evaluation set created by Kakao is plausible, but the result is not independent validation. The 72.50 score should be treated as a company-reported in-house result until an external evaluation reproduces it. The card labels the set as in-house, to its credit. Readers should keep the label attached to the number.
The speed number
Korean coverage of the launch reports Kakao's claim that average response time stayed below 0.04 seconds on a single Nvidia H200. That figure does not appear on the model card, and Kakao has not published a test method alongside it. No prompt set, batch conditions or measurement protocol are provided in the material reviewed for this article.
What is usable is narrower: a small dedicated decider is cheap to run per judgment, and anyone evaluating it can reproduce the latency test on their own workload in an afternoon, because the weights are public.
Sources
- [1] Kakao, Kanana-2-v-decide-3b-2611 model card, Hugging Face (release note dated November 6, 2026; architecture; the choice, noul and score question types; input limits; the full Decision Index 0.4.1 and Kanana Decision Bench table; the one-request-at-a-time serving note)Read source
- [2] Kakao, Kanana Open License Agreement, license text in the model repository (royalty-free use and derivative works; Section 4.2 permits use in your own services without a separate commercial license; Section 4.1 reserves resold access, system-integration and on-premise resale and on-device embedding for a separate Kakao commercial license; Section 3.1 redistribution conditions include a notice file and a "Powered by Kanana" display)Read source
- [3] Seoul Economic Daily, "Kakao Releases AI Decision Model for Faster Agent Judgments", November 6, 2026 (Kakao's statement that average response time stayed within 0.04 seconds on a single Nvidia H200; Kakao's plan to apply the model first to web search and automatic classification of product categories)Read source