Kakao has released an open-weight AI model that answers questions with probabilities, choices and scores instead of sentences, aiming to make the rapid judgments AI agents need while they work. The model, Kanana-2-v-decide, was published on Hugging Face on 6 November under a licence that permits commercial use. Kakao says it answers in under 0.04 seconds on a single Nvidia H200 graphics card, about 43 times faster than GPT-6 Luna in its own comparison.
## A model that decides instead of chatting
Kanana-2-v-decide does not generate text. Given a situation — text, a JSON object, a chat history, or images — it answers questions: pick one option from a list, answer yes or no, or rate something on an ordered scale. It returns a probability for each possible answer, so an agent can act on the most likely one without parsing text.
The model has 3.67 billion parameters and is built on Kakao’s own multimodal model, Kanana-2-v-3b. It accepts up to 10 images and 20,480 tokens of input, and can answer any number of questions in a single request. Kakao plans to use it first inside its own services, including web search and automatic classification of product categories.
Developers can run the model through the transformers library or serve it over HTTP through Kakao’s System One API, which exposes the same questions-and-answers interface. The model needs a CUDA graphics card and the transformers library, and Kakao publishes the full serving code alongside the weights.
## The speed claim — and the fine print
The headline figure is speed. Kakao says the model answers in under 0.04 seconds on one H200 card, about 43 times faster than GPT-6 Luna, the model it used for comparison. That figure comes from Kakao’s own tests, not an independent benchmark, and the comparison pits a small specialist against a general-purpose model.
The honest reading is narrower: for yes-or-no and multiple-choice judgments, a small specialist can be far faster than a big generalist. The benchmark scores tell a similar story. The model leads on Decision Index Vision and averages 61.23 across two public decision benchmarks, but it does not top every column — a rival leads on the text-only Decision Index.
On Kakao’s own Korean-language test, Kanana Decision Bench, the model scored 72.50 across 15,909 questions — 18 points ahead of the next-best model compared. That test is Kakao’s own creation, so the public benchmarks matter more: there it averages 61.23, leading on the vision test while a rival leads on the text-only test.
## Our Read
Decision models are a quiet but important shift in how agents get built. Instead of asking a large language model to write an answer and then parsing the text, a developer can ask a small model for a probability and act on it directly — faster, cheaper and more predictable.
Kakao is not alone in this. Open-weight releases this week include [Kimi K3.1 Pro](https://geroundtable.com/kimi-k3-1-pro-open-weights/), while frontier launches like [Inkling 2 Small](https://geroundtable.com/thinking-machines-inkling-2-small/) show the other end of the scale. OpenAI’s GPT-6 line keeps expanding too, with [GPT-6 Cyber](https://geroundtable.com/gpt-6-cyber-critical-security-model/) restricted to vetted security customers. Kakao’s bet is that the judgment layer, not the chat layer, is where agents will spend their time.
## Why It Matters
If decision models catch on, the agent stack splits in two: big models that reason and write, and small models that judge quickly and cheaply. That matters for cost — a 0.04-second judgment on one graphics card is far cheaper than a full model call — and for reliability, since a probability is easier to audit than a paragraph.
Kakao’s release is a small model, but it points at a bigger change in how agents will be built. The first deployments will be inside Kakao’s own services, where the company can prove the speed and accuracy claims in production before outside developers bet on them.
## Background
Kakao is South Korea’s dominant messaging and platform company, and its Kanana model family is its bid to stay relevant as AI agents reshape its services. The company released two Kanana-2 embedding models as open source on 29 October, and Kanana-2-v-decide follows a week later. The model carries the Kanana Open License, which permits research and commercial use.
Source: [Kakao’s model card](https://huggingface.co/kakaocorp/kanana-2-v-decide-3b-2611) and [Seoul Economic Daily’s report](https://en.sedaily.com/technology/2026/11/06/kakao-releases-ai-decision-model-for-faster-agent-judgments)
0 Comments