Open weights
FreeDownload; Kanana Open LicenseUse, modify, fine-tune and redistribute, including in your own services. Redistribution must show "Powered by Kanana" and keep the license notice. Official source 1 ↗ · Official source 2 ↗
Reviewed tool profile
Kanana-2-v-decide
Kakao's 3B multimodal decision model that returns probabilities instead of text.
Sources checked
- Pricing
- The weights are free to download and use under the Kanana Open License, including in your own services. Reselling access to the model or its derivatives through an API or cloud service, as part of a system-integration or on-premise solution, or embedded in on-device products needs a separate commercial license from Kakao, granted at Kakao's discretion. Kakao releases it as weights you run yourself. Official source 1 ↗ · Official source 2 ↗
- Available as
- Self-hosted from the Hugging Face weights with Hugging Face Transformers (custom code), in Python directly or through a pipeline. · A bundled HTTP server script exposes a Jev-compatible System One API, so existing TypeSafe SDK clients can call it by changing the base URL. Official source 1 ↗ · Official source 2 ↗ Official source 1 ↗ · Official source 2 ↗
- Primary group
- Automation & Agents
Also consider
See all 6 similar tools ↓- Claude Haiku 5.5Anthropic’s fastest, lowest-cost Claude model for high-volume tasks and coding subagents.Free plan
- ifyAI support agent that resolves tickets inside Freshdesk, Zendesk, Salesforce or HubSpot, billed per verified resolution.Free trial
- LFM3-1.7BLiquid AI's 1.7B text model for local agents that fits in about 1 GB on a phone or laptop.Free plan
What it does
Kanana-2-v-decide is Kakao's open-weight decision model, released on Hugging Face on 6 November 2026. Instead of writing text, it answers typed questions about an input — pick an option, yes or no, or a score on a scale — and returns a probability for every possible answer. The input can be text, JSON or a chat history with up to 10 images, so it suits the repeated small judgements inside AI agents and services: routing a customer inquiry, checking a product photo, or choosing which tool to call next. It has 3.67B parameters and is built on Kakao's Kanana-2-v-3b multimodal model.
Official source 1 ↗ · Official source 2 ↗Features
- Three question types, probabilities out. Choice picks one of several options, noul gives the probability a statement is true, and score rates on an ordered 2–10 level scale; every answer comes with a full probability distribution and a confidence value. Official source 1 ↗
- Text and images together. The state can be plain text, a JSON object or array, or chat messages, with up to 10 images and 20,480 tokens per request. Official source 1 ↗
- Many questions in one call. One request can carry any number of questions — Kakao's example reads a damaged-delivery complaint and its photo and decides damage, inquiry type and customer sentiment at once. Official source 1 ↗ · Official source 2 ↗
- Unlimited, order-free options. Options are embedded separately and compared with the question, so there is no fixed cap on the number of choices, option order does not sway the result, and repeated option lists are cached. Official source 1 ↗ · Official source 2 ↗
- Calibrated confidence. Kakao calibrated the confidence scores on validation data, cutting calibration error from 2.06% to 1.12%, so low-confidence decisions can be thresholded and sent for review. Official source 1 ↗
- Fast per decision. Kakao reports about 42 ms per decision on one H200 GPU on Korean product-review data, which it says is about 43 times faster than GPT-6 Luna; latency varies with token count. Official source 1 ↗
- Leads small decision models on Kakao's tables. Kakao's model card shows 72.50 on its Korean Kanana Decision Bench and a 61.23 public-benchmark average, ahead of StartLux-Decision, Liquid AI d1-3B and Base Labs Joey-1; its blog puts it slightly ahead of Microsoft-Decision-1 and Jev on Korean text. Official source 1 ↗ · Official source 2 ↗
Pricing & free limits
Commercial license
Contact KakaoCustomRequired to resell access through an API or cloud service, in SI or on-premise solutions, or embedded on devices; granted at Kakao's discretion. Official source 1 ↗
Billing details
The license is royalty-free: Kakao charges nothing for the weights, and it claims no rights to the outputs you generate. Official source 1 ↗
Free access & limits
No usage limits on the weights themselves; you pay only for your own hardware. Official source 1 ↗
Best for & limitations
Routing customer inquiries by type, checking for damage in an attached photo and scoring sentiment in one call. Official source 1 ↗
Letting an AI agent choose which tool to call next, as a fast step between larger-model calls. Official source 1 ↗
Checking whether a product photo matches a search query or violates a guideline, and classifying products into categories. Official source 1 ↗
Korean-language moderation and classification pipelines that need probabilities they can threshold. Official source 1 ↗ · Official source 2 ↗
- Not open source in the OSI sense. Despite Kakao calling it open source, the Kanana Open License requires a separate commercial license to resell API, on-premise or on-device access, and models trained on it or its outputs must carry a "Kanana" name prefix. Official source 1 ↗
- Does not generate text. It only returns probabilities over the options you supply, and those probabilities shift when you add or change an option. Official source 1 ↗
- Korean and English only. The model card lists Korean and English; Kakao's own benchmark is a Korean evaluation set. Official source 1 ↗
- Vendor-reported speed and scores. The 42 ms latency and the benchmark tables come from Kakao, including an in-house Korean evaluation set. Official source 1 ↗ · Official source 2 ↗
- Input limits. No video input, extreme aspect ratios are not supported, EXIF orientation is ignored, and overlong text is silently shortened by default. Official source 1 ↗
- Simple bundled server. The included HTTP server handles one request at a time and must be given an API key before you expose it beyond localhost; new option lists also need a first embedding pass. Official source 1 ↗ · Official source 2 ↗
Similar tools
Tools like Kanana-2-v-decide
Published tools that share its categories.
Claude Haiku 5.5Anthropic’s fastest, lowest-cost Claude model for high-volume tasks and coding subagents.
Visit ifyAI support agent that resolves tickets inside Freshdesk, Zendesk, Salesforce or HubSpot, billed per verified resolution.
Visit LFM3-1.7BLiquid AI's 1.7B text model for local agents that fits in about 1 GB on a phone or laptop.
Visit 
GlasshandRent a real iPhone or Mac that your AI agent controls through an API.
Visit 
FastinoOpen-weight decision models, led by GLiDE, that return probabilities over your own outcomes for agent routing, classification and scoring.
Visit 
RerunNo-code AI coworkers that run your recurring business tasks on a schedule or trigger, on your own ChatGPT or Claude plan.
Visit