Kakao releases open AI model that reads text and images to make decisions

Kakao has released Kanana-2-v-decide, an AI model that reads text and images and returns a decision rather than a written reply. Kakao built it to classify customer inquiries and help AI agents pick tools. The company published it on Hugging Face on Nov. 6 under its own Kanana Open License, which allows research and commercial use.

The key design choice is the kind of output. Chat models answer in sentences. This model answers in probabilities, selections and scores, which software can act on directly. For builders wiring AI into support queues or agent pipelines, that makes it closer to a fast classifier than to a chatbot.

What Kakao released on Hugging Face

Kanana-2-v-decide is an internally developed multimodal model, meaning it takes more than one kind of input. It is built on Kakao’s own multimodal language model, Kanana-2-v-3b, and accepts both text and images. The Elec’s report does not give a parameter count for either model, and Kakao has not published accuracy benchmarks in the available coverage.

The model has three output modes:

• Condition check: it calculates the probability that a specific condition is true, for example “this product is damaged”.
• Option choice: it assigns probabilities to several options and picks the most suitable one, which fits routing and tool selection.
• Graded score: it produces a score that reflects the probabilities of different evaluation levels.

Developers choose the output format their service needs and pass the result to the next step in their pipeline.

Spec	Kanana-2-v-decide (as reported)
Maker	Kakao
Release	Nov. 6, 2026, on Hugging Face
Base model	Kanana-2-v-3b, Kakao’s multimodal language model
Inputs	Text and images
Outputs	Probability of a condition, option choice with probabilities, or graded score
Architecture	Embedding-based comparison of inputs and candidate options
Speed	Average under 0.04 seconds on one Nvidia H200 GPU, per Kakao
Licence	Kanana Open License, research and commercial use permitted
Parameter count	Not stated in the report

How a decision model differs from a chat model

Most multimodal models in production generate text. Asked whether a photo shows a broken item, they reply with a sentence, and the developer then has to parse that sentence into a yes, a no, or a category. Kanana-2-v-decide skips that step by returning the decision itself, expressed as probabilities or scores.

Kakao’s own example is a customer who sends a photo of a damaged product along with a replacement request. The model can analyse the image and the message to decide whether the product is damaged, classify the inquiry, and assess the customer’s emotional state. The service then applies predefined rules to route the customer to a support channel or send replacement instructions.

The pattern matters. The model does not decide what the company does. It produces structured signals, and the business logic that acts on them stays in ordinary, auditable rules. That separation is useful in support and commerce workflows, where a wrong free-text answer is harder to catch than a wrong label.

Tool selection for agents

Kakao also positions the model for AI agents, meaning systems that carry out multi-step tasks and call external tools such as search, booking or payment functions. In the option-choice mode, the candidate tools are the options and the model scores which one fits the request. The report does not describe how many options the model handles well, or how it performs against general-purpose models that do tool calling through generated text.

Why Kakao chose an embedding design for speed

Under the hood, Kakao uses an embedding architecture. An embedding converts an input, such as a message plus a photo, into a list of numbers called a vector. Similar meanings end up as nearby vectors. Kanana-2-v-decide converts both the input and each candidate option into vectors and compares how similar they are.

The practical gain is reuse. Vectors that represent the same options can be stored and used again, so the model does not recompute them on every request. Kakao says this reduces computational load and response latency, the delay before an answer arrives, when the same kinds of requests repeat. A support system with a fixed set of categories, or an agent with a stable toolbox, fits this pattern.

Kakao reports an average response time of less than 0.04 seconds in tests on a single Nvidia H200 graphics processing unit (GPU). That is the company’s own measurement. The report does not give the input sizes, image resolutions, batch settings or number of options behind it, so it is a rough indication rather than a figure to plan capacity around.

For a sense of scale within Kakao’s own line-up, the company said in 2024 that its earlier multimodal model Kanana-o produced answers in an average of 1.6 seconds regardless of the question. The two models do different jobs, generating answers versus scoring options, so the numbers are not a like-for-like comparison. They do show where a decision-only design can save time.

Access, licence and Kakao’s own deployment plans

The model is available on Hugging Face, the public hub for model files. The Elec describes the release as open source, citing Kakao. The licence, though, is the Kanana Open License, Kakao’s own terms, which the report says permit research and commercial use. Custom licences often add conditions that standard open-source licences do not, so teams should read the full text before shipping a product on it. The report does not mention a hosted API from Kakao.

Kakao plans to use the model first in internal projects, including web search and automatic product category classification, to improve its performance. After that, the company intends to combine search and decision-making in a single model.

Where the model fits in Kakao’s Kanana and agent plans

Kanana is the brand Kakao gave its in-house model family when it unveiled ten models at its if kakao 2024 developer event: three large language models, three multimodal language models, two image and video generation models, and two speech models. At the time Kakao said it was aiming for models tuned to its services and cost efficiency rather than the largest possible size. Kakao also said it would combine its own models with external and open-source models depending on the service, choosing the cheapest option where performance is equal. A small, fast decision model fits that approach.

The release also follows Kakao’s stated push into agents. On its second-quarter earnings call in August 2025, CEO Chung Shin-a said the company was preparing an AI agent ecosystem and planned to release inference models that are essential for AI agent services. Around the same time, Korean coverage described Kakao’s agent work as including the Kanana assistant app, “AI Mate” services, and an agent being developed with OpenAI for tasks such as restaurant reservations, taxi calls and payments inside KakaoTalk. Choosing the right tool for each request is the kind of step such agents need, and that is the job Kakao describes for Kanana-2-v-decide.

In October 2026, Kakao also said it would extend its AI safety evaluations from models to AI agents. The available coverage does not say whether Kanana-2-v-decide has gone through those evaluations.

What it means for builders, and the open questions

For developers building support desks, moderation queues, commerce back-ends or agent routers, the release offers a few concrete things:

• Structured output by design. Probabilities and scores feed straight into thresholds and rules, with no prompt engineering needed to force a model into a fixed format.
• Images as first-class input. Photo-plus-text cases, such as damage claims or product listings, can be handled in one call.
• Cacheable options. Stable categories or tool lists can be embedded once and reused, which Kakao says cuts compute and latency.
• Commercial use permitted under the Kanana Open License, according to the report.

The limits are just as clear from what the sources do not say:

• No accuracy figures. The only reported number is latency. There is no public evidence yet on how well the model classifies inquiries or picks tools.
• Size not stated. The report names the base model, Kanana-2-v-3b, but gives no parameter count for the decision model, which affects hardware planning.
• Languages not stated. The coverage does not say which languages the model is tuned for, which matters for teams outside Korea.
• Thin test details. The 0.04-second figure comes without information on workload, batch size or image size.
• Sensitive signals. Kakao’s example includes assessing a customer’s emotional state. Teams using that kind of output to route people should consider how errors and bias would affect customers, a concern that agent-level safety reviews like Kakao’s own are meant to cover.

Probability outputs also need calibration in practice. A score of 0.8 is only useful if it means roughly what it says on the team’s own data. Any team adopting the model will need to test it on its own inquiries before trusting thresholds.

What to watch next

Kakao’s next steps are already sketched out. Watch for results from its internal use of the model in web search and product categorisation, and for the planned single model that combines search with decision-making. Independent tests that measure classification and tool-selection accuracy, and that check the latency claim on other hardware, will show whether the design holds up outside Kakao. It is also worth watching how the model features in Kakao’s agent services and the agent-level safety reviews the company has committed to.

For other recent releases built around AI agents, see Perplexity’s Teammates and Tencent’s Hy4, which can see images and use a computer.