Kakao Releases Kanana-2-v-decide Weights to Support Agent Tool Selection with Multimodal Decisions

On November 6, Kakao announced the release of its multimodal decision model, Kanana-2-v-decide. The official Hugging Face collection lists kanana-2-v-decide-3b-2611. The release targets recurring classification, tool selection, and image judgment tasks in agent workflows, allowing services to obtain results that can be processed directly by software. The weights are under the Kanana Open License. Kakao says research and commercial use are permitted, but the license terms should still be reviewed before deployment. Official announcement, model collection

Based on Kakao’s Kanana-2-v-3b, the model offers three decision formats: Noul returns the probability that a condition is met; Choice returns option probabilities and the selected answer; and Score calculates a score from probabilities at each rating level. A single request can include multiple questions. For example, after receiving a product photo and a return or exchange message, it can determine whether the item is damaged, identify the inquiry type, and assess the customer’s sentiment, then leave the next action to business rules. Technical overview

The architecture represents questions and candidate answers as vectors, then compares their similarity. Vectors for fixed options can be stored in advance and reused, reducing repeated computation. From an engineering perspective, this could make it suitable to extract high-frequency decisions with clearly defined answer sets into a standalone service, reducing the cost of having an agent generate text at every step and then parse the answer. The actual benefit will depend on input length, the number of candidates, and cache hit rates. Architecture overview

The reported speed figures need to be read in light of their test conditions. The technical article reports about 42 milliseconds per decision on a single H200 using Korean product review data, and notes that latency varies with token count. The press release summarizes the result as an average of under 0.04 seconds. The figures differ slightly, and both come from vendor tests; they should not be treated as latency guarantees for Chinese, multiple images, or concurrent serving. Test conditions

For Chinese-language teams, the next step is to measure classification accuracy, probability calibration, and tail latency using their actual tool lists and image data, and to validate how low-confidence results are handed off to another model or a human. These deployment metrics will determine whether the model can reliably handle decision steps in an agent workflow.