---
type: "post"
id: "kakao-open-weights-kanana-2-v-decide-multimodal-d8fb70"
slug: "kakao-open-weights-kanana-2-v-decide-multimodal-d8fb70"
title: "Kakao open-weights Kanana-2-v-decide multimodal decision model (~3.67B)"
description: "Kakao has released Kanana-2-v-decide, an open-source multimodal model that returns probabilities, choices or scores instead of generated prose."
retrieval_nugget: "Kakao has released Kanana-2-v-decide, an open-source multimodal model that returns probabilities, choices or scores instead of generated prose. The company positions it for recurring decisions inside AI services and agents, including request routing, tool selection and image judgments. Its reported speed could make these steps easier to run synchronously, though the performance claims need independent testing."
published_at: "2026-11-06"
updated_at: "2026-11-06"
record_date: "2026-11-06"
date_kind: "published_at"
topics: ["multimodal"]
entities: ["Kakao","Kanana-2-v-decide","Kanana-2-v-3b","Hugging Face","Kanana Open License","TypeSafe AI","Microsoft","Perplexity"]
related_patterns: ["model-answers-become-task-interfaces"]
novelty_label: "new_fact"
what_changed: "Kakao released Kanana-2-v-decide, an open-weights multimodal decision model with structured probability, choice and score outputs, reporting sub-0.04-second latency and commercial-use licensing."
prior_refs: ["https://newruntime.com/posts/vllm-semantic-router-releases-d3-ultra-a-26b-37f577/","https://newruntime.com/posts/open-d1-edge-decision-models-for-text-vision-e8468e/"]
source_url: "https://www.kakaocorp.com/page/detail/12193"
source_urls: ["https://www.kakaocorp.com/page/detail/12193"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:kakao-open-weights-kanana-2-v-decide-multimodal-d8fb70"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/kakao-open-weights-kanana-2-v-decide-multimodal-d8fb70/","markdown":"https://newruntime.com/posts/kakao-open-weights-kanana-2-v-decide-multimodal-d8fb70.md","json":"https://newruntime.com/posts/kakao-open-weights-kanana-2-v-decide-multimodal-d8fb70.json"}
markdown_url: "https://newruntime.com/posts/kakao-open-weights-kanana-2-v-decide-multimodal-d8fb70.md"
---

# Kakao open-weights Kanana-2-v-decide multimodal decision model (~3.67B)

## Retrieval answer

Kakao has released Kanana-2-v-decide, an open-source multimodal model that returns probabilities, choices or scores instead of generated prose. The company positions it for recurring decisions inside AI services and agents, including request routing, tool selection and image judgments. Its reported speed could make these steps easier to run synchronously, though the performance claims need independent testing.

## Summary
Kakao has released Kanana-2-v-decide, an open-source multimodal model that returns probabilities, choices or scores instead of generated prose. The company positions it for recurring decisions inside AI services and agents, including request routing, tool selection and image judgments. Its reported speed could make these steps easier to run synchronously, though the performance claims need independent testing.

## What happened
Kakao announced on November 6, 2026, that it had released Kanana-2-v-decide on Hugging Face. The company says it developed the model from its Kanana-2-v-3b multimodal language model. Rather than produce a natural-language answer, it returns a structured decision in one of three formats: probability, choice or score. A probability estimates whether a condition is true; a choice provides probabilities for options and identifies the preferred one; and a score provides a more granular assessment based on probabilities at each evaluation stage.

The model is intended for decisions that can feed directly into application logic. Kakao gives the example of a customer asking to exchange a damaged item and attaching a photo. The model can assess the image and inquiry together, estimating whether the item is damaged, what kind of inquiry it is and the customer’s emotional state. A service could then combine those outputs with its own rules to route the case or explain the exchange process. The company also names agent tool selection and product-image judgment as potential uses.

Kakao says the model uses an embedding-based structure: questions and options are represented as vectors and compared for similarity. The company says this supports flexibility when there are many classification categories, and that vectors for repeated options can be stored and reused to reduce computation and response delay on recurring requests. On one H200 GPU, Kakao reports an average response time below 0.04 seconds, which it describes as 43 times faster than GPT-6 Luna. These figures are Kakao’s measurements.

For quality, Kakao reports the highest performance among the models it compared on its Korean-language Kanana Decision Bench, comprising 15,909 questions, and on the public Decision Index dataset. The comparison set included Microsoft Decision-1, TypeSafe AI’s Jev, Perplexity’s pplx-decider-v1.1-27B and Upstage’s Solar Decide. Kakao says the release uses its Kanana Open License, which permits research and commercial use. The company plans first to apply it internally to web search and automatic product-category classification, then continue improving it in service settings. Its stated longer-term direction is to combine search and decision-making in one model.

## What the archive adds
New Runtime’s earlier coverage described open d1 as an open-weight decision-model release for text, vision and audio, with edge-latency benchmarks and Hugging Face availability ([Open d1: Edge decision models for text, vision, and audio](https://newruntime.com/posts/open-d1-edge-decision-models-for-text-vision-e8468e/), October 24, 2026). Kanana-2-v-decide is a new release and another example of teams publishing specialized decision models; it is not the same model or release.

## Trend signal
This supports the pattern that **model answers become task interfaces**: outputs are shifting from general prose toward bounded formats assembled for a specific task ([Model answers become task interfaces](https://newruntime.com/patterns/model-answers-become-task-interfaces/)). Kanana’s probability, choice and score outputs are intended to plug into application logic without first extracting a decision from generated text.

## Why now
Kakao frames the release around the repeated decisions agents need to make as they select tools, consult information and choose subsequent steps. It also says it will test the model first in internal search and product-classification projects, linking the release to a stated effort to improve the model in actual service environments.

## Why it matters
For teams building agents, many steps are not open-ended conversation: they are bounded questions such as which tool to call, how to route an inquiry or whether an image meets a criterion. A model that returns a structured result directly may remove a parsing step and make it easier to connect decisions to deterministic application rules. Kakao’s vector-reuse description also points to a possible efficiency benefit when the same options recur, although the release does not quantify that benefit separately.

The practical case is strongest where a team has repeatable choices and can define the allowed outputs. Builders should treat Kakao’s speed and benchmark claims as hypotheses, not guarantees: measure latency, accuracy and failure costs against their own inputs and baselines. The commercial license may make experimentation easier, but production adoption still depends on checking the license terms and model behavior for the intended use.

## Limits and uncertainty
The announcement reports company-run performance results but does not establish independent replication of its speed or benchmark rankings. The 43× comparison is Kakao’s stated comparison with GPT-6 Luna; it does not by itself show equivalent hardware, workload or serving conditions. The release also does not quantify error rates on deployed customer requests, the impact of vector reuse, or how model decisions should be calibrated for high-consequence workflows. Teams will need to test those questions on representative data before relying on the outputs.

## Reader takeaway
Kanana-2-v-decide is a commercially usable, open-licensed model for turning text and images into structured decisions rather than prose. It is worth evaluating for repeated classification and choice steps in agents, but Kakao’s headline performance figures should be validated against a team’s own workloads before they shape a production design.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Multimodal](https://newruntime.com/topics/multimodal.md): Explore the multimodal topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)