Kakao Releases Open-Weight Decision Model Kanana-2-v-decide-3b
The 3.67B-parameter model outputs probability distributions over predefined answers instead of text, targeting low-latency classification and routing in agent pipelines.
Kakao has released Kanana-2-v-decide-3b-2611, an open-weight multimodal model designed for decision tasks rather than text generation. Published on Hugging Face on 6 November 2026, the model accepts a state consisting of text, JSON objects, chat history, and up to ten images within a 20,480-token context. Instead of producing free-form output, it returns a probability distribution over predefined answers, supporting choice, yes/no, and ordinal scoring question types.
Architecturally, the model encodes the state and each candidate answer into 2,560-dimensional embeddings and applies a softmax over cosine similarities, allowing multiple questions to be answered in a single request. Vendor-reported results include a combined Decision Index average of 61.23 and latency below 0.04 seconds on a single NVIDIA H200, which Kakao describes as roughly 43 times faster than a generic LLM for decision tasks.
The 3.67-billion-parameter model supports Korean and English and is distributed under the Kanana Open License, which permits research and commercial use subject to notification, attribution, and responsible-AI guideline conditions.