Kakao open-sources Kanana-2-v-decide, a multimodal model that outputs probabilities, choices and scores for agents
Kakao says the model answers in under 0.04 seconds on one NVIDIA H200 and about 43 times faster than GPT-6 Luna, by its own measurement; the benchmarks are Kakao's and unaudited.
Why it matters
Agents repeatedly decide which tool to call, which source to trust and what to do next, and today they usually pay a general language model to write a paragraph for each choice. Kakao's release targets that loop with a model that returns a structured probability, selection or score instead. The speed and benchmark figures come from Kakao's own Kanana Decision Bench and the public Decision Index, so independent tests are the thing to watch, along with whether other labs copy the narrow decision-model format.
Key facts
- Open-sourced on Hugging Face on 6 November
- Average response time under 0.04 seconds on one NVIDIA H200
- About 43 times faster than GPT-6 Luna by Kakao's own measurement
- Kanana Decision Bench: 72.62 vs 71.98 for Microsoft Decision-1
- Decision Index: 61.35 for the 3B model