Kakao has released a small AI model that does not write sentences at all — it makes decisions. Kanana-2-v-decide answers an AI agent’s questions in probabilities, choices and scores, and the company says it does so in under 0.04 seconds.

Key Points

  • Kakao released Kanana-2-v-decide, a multimodal decision-making model, as open source on Hugging Face on Friday, 6 November.
  • Instead of generating text, the model returns decisions in three formats — probability, choice and score — so AI agents can classify requests and pick tools.
  • Built on Kakao’s Kanana-2-v-3b model, it reads text and images together and has 3.67 billion parameters, according to its Hugging Face page.
  • Kakao says it averaged under 0.04 seconds per response on a single Nvidia H200 GPU, about 43 times faster than GPT-6 Luna in the company’s own comparison.
  • It is released under the Kanana Open License, which permits research and commercial use.
  • Kakao plans to use it first in its own web search and automatic product-category classification.

Kakao released a new AI model on Friday, 6 November, built for a job most language models are not designed to do: making fast, structured decisions for AI agents, rather than writing replies.

What happened

The model, called Kanana-2-v-decide, was published on Hugging Face, the global platform for sharing AI models, under a licence that allows both research and commercial use. It is built on Kanana-2-v-3b, Kakao’s own multimodal language model, and it understands text and images together.

Where a conventional chatbot answers a question with a paragraph, Kanana-2-v-decide answers with a decision. It can calculate the probability that a given condition is true, pick the most suitable option from a set of candidates, or return a score on an ordered scale. It does not generate text. That makes it suited to the repetitive judgements an AI agent must make while carrying out a task — sorting a customer inquiry, choosing which tool to use next, or judging a product image — without waiting for a full language model to compose a reply.

Speed is the point. Kakao says the model recorded an average response time of less than 0.04 seconds running on a single Nvidia H200 graphics processor. The company says that is about 43 times faster than GPT-6 Luna, based on its own measurements. The model uses an embedding-based design: questions and answer options are turned into vectors, numerical groupings, and compared for similarity, rather than being processed as sentences.

On benchmarks, Kakao reports the model scored highest among the systems it compared on both its own Korean-language evaluation set, Kanana Decision Bench — 15,909 questions, on which it scored 72.50 — and the public Decision Index benchmark, where its Hugging Face page reports an average of 61.23. Rivals named in the comparison included Microsoft’s Decision-1, TypeSafe AI’s Zeb, Perplexity’s pplx-decider-v1.1-27B and Upstage’s Solar Decide. Those results are Kakao’s own figures, not independent tests.

“Kanana-2-v-decide provides decisions in standardised formats so AI agents can act quickly and consistently,” said Choi Dong-jin, leader of Kakao’s Applied AI Model Performance team.

Why it matters

The release lands in the middle of a shift in how AI is built. The newest systems are not single chatbots but agents — software that plans, picks tools and carries out multi-step work on a user’s behalf. Every step of that work needs a judgement: which tool, which option, how confident. A model that answers those questions in milliseconds, and in a format a program can act on directly, removes a bottleneck that has slowed agents down.

Kakao is not keeping the model to itself. Publishing the weights openly, with a licence that allows commercial use, puts the tool in the hands of outside developers and researchers, who can build their own agent services on it. The company released two embedding models, Kanana-2-embedding, as open source on 29 October, and the decision model extends the same open strategy.

The timing also matters for Kakao itself. On a 5 November earnings call, chief executive Chung Shin-a said the company will begin piloting an integrated AI workspace called Kanana Space in December. A fast decision model is the kind of component that sits underneath such a product — classifying what a user wants, routing it to the right tool, and doing it fast enough that the user never notices the wait.

The week has been busy for open AI releases. Google’s Gemini Omni 1.2 Flash arrived with longer, cheaper video generation, and Tencent’s Hy4 open-source model launched two days earlier, underlining how quickly the open-model race is moving.

What happens next

Kakao says it will deploy Kanana-2-v-decide first in its own services, starting with web search and automatic classification of product categories, and refine it in live use before it spreads more widely. Independent testing will show whether the benchmark lead holds outside Kakao’s own evaluation sets — the 43-times-faster figure, in particular, is the company’s own comparison and has not been verified independently.

For developers, the model is available to download now from Hugging Face. What to watch is whether other agent builders adopt it, and whether the decision-model format — probabilities, choices and scores instead of sentences — becomes a standard building block for the next generation of AI agents.

Sources: Seoul Economic Daily, 6 November 2026; The Elec, 6 November 2026; Aju Press, 6 November 2026; eDaily, 6 November 2026; Herald Business, 6 November 2026; Asia Today, 6 November 2026; Hugging Face model card (kakaocorp/kanana-2-v-decide-3b-2611). The claim that Kanana-2-v-decide is about 43 times faster than GPT-6 Luna, and the benchmark scores, are Kakao’s own figures and are labelled as company claims in the article. Published: 6 November 2026, Kuala Lumpur time.