Kakao's New AI Model Refuses to Write Sentences. It Just Decides.

A model that answers without writing a word

Most AI models answer you with paragraphs. Kakao's newest one refuses. Kanana-2-v-decide, released open source on Hugging Face on November 6, is an AI decision model that takes text and images as input and returns a decision: a probability, a choice between options, or a score on a scale. No sentences, no small talk, no paragraphs to parse.

That sounds like a small difference. It is not. Inside an AI agent, the moment a model writes a sentence, something else has to read that sentence and figure out what it decided. A probability or a score skips that step entirely. The decision is the output, ready for code to act on.

Kakao built the model on its own Kanana-2-v-3b multimodal language model. It has 3.67 billion parameters, with a 0.49 billion parameter vision encoder paired with a 3.18 billion parameter language model. It accepts up to 10 images and 20,480 tokens per request, and it handles three kinds of questions: yes or no, pick one of several options, or rate something on an ordered scale.

Fast enough to disappear

Speed is the whole point of a model like this. Kakao says Kanana-2-v-decide averages under 0.04 seconds per response on a single Nvidia H200 GPU. The company claims that is 43 times faster than GPT-6 Luna on the same kind of task, though that comparison comes from Kakao's own testing and should be read as a vendor claim until independent benchmarks weigh in.

On accuracy, Kakao reports the model scored 72.50 on its own Korean-language Kanana Decision Bench, 18 points ahead of the next system it tested, and averaged 61.23 across the public Decision Index and Decision Index Vision benchmarks. The competitors it names include Microsoft's Decision-1, TypeSafe AI's Zeb, Perplexity's pplx-decider-v1.1-27B, and Upstage's Solar Decide. Again, these are Kakao's numbers on benchmarks it partly built itself, so treat them as a starting point, not a verdict.

The license matters as much as the speed. Kanana-2-v-decide ships under the Kanana Open License, which permits research and commercial use. That puts it in the hands of any developer building agents, not just Kakao's own teams.

Where a decision model actually earns its keep

The obvious uses are the boring ones, and boring is good here. Sorting customer inquiries, deciding which tool an AI agent should call next, checking whether a product photo matches its listing, grading an answer against a rubric. These are judgments that software already makes, usually with brittle rules or by asking a giant language model to write an essay and then parsing the essay for a verdict.

A decision model replaces both. It is small enough to run cheaply, fast enough to sit inside a live workflow, and its output needs no interpretation. Kakao plans to deploy it first inside its own services, starting with web search and automatic product category classification.

That fits a wider pattern this year. The frontier labs keep shipping bigger models, but a parallel race is running in the opposite direction: small, specialized models built to do one job well. Meituan's LongCat-2.5 targets long-horizon agent work, Tencent's Hy4 adds vision and computer use to an open-weight model, and now Kakao is betting that the cheapest way to make agents reliable is to stop asking them to write.

The catch: decisions need trust, not just speed

A model that outputs a number instead of a sentence is easier to use, but it is also easier to trust blindly. A probability of 0.92 looks precise. It is still a guess, and when the model is wrong, there is no reasoning to inspect, no chain of thought to audit. Kakao says the model returns a full probability distribution over the supplied options, which at least shows how close the call was. Developers should use that uncertainty, not hide it.

There is also the question of what the model was trained to decide. A decision model inherits the biases of its training data just like any other model, and a biased score can quietly gate real outcomes: which support ticket gets priority, which product gets flagged, which applicant gets routed where. The speed makes those calls cheaper. It does not make them fairer.

What to try first

If you build agents, this is worth an afternoon. The model is on Hugging Face as kakaocorp/kanana-2-v-decide-3b-2611, and Kakao ships a small HTTP server script with it. The practical test is simple: take one place in your workflow where you currently ask a large model to classify something or pick a tool, swap in the decision model, and compare cost, latency, and accuracy over a few hundred real examples.

Kakao is not stopping here. The company said on its November 5 earnings call that it will begin piloting Kanana Space, an integrated AI execution space inside KakaoTalk, in December, and formally launch two PC services, Kanvas and Modu-ui AI, before the end of the year. The decision model is a building block for that larger push, and its open release is a bet that developers will find uses Kakao has not thought of yet.

The AI agent wars have mostly been fought over who can write the best prose. Kakao just argued, with a model release, that the next fight is over who can decide fastest. It is a good argument. The benchmarks will tell us whether the model lives up to it.