Kanana-2-v-decide

Kanana-2-v-decide is Kakao's multimodal decision model, open-sourced 2026-11-06. Instead of free-form text, it returns structured probability / choice / score judgments so agents can route tools, triage tickets or score product images quickly. Press: https://www.kakaocorp.com/page/detail/12193 ; HF collection: https://huggingface.co/collections/kakaocorp/kanana-2-v-decide ; weights e.g. kakaocorp/kanana-2-v-decide-3b-2611.

Highlights (official):

• Built on multimodal Kanana-2-v-3b; text + image inputs;
• Embedding-similarity design that can cache choice vectors for large label sets;
• ~0.04 s mean latency on one H200 — claimed ~43× faster than GPT-6 Luna;
• Tops Kakao's Kanana Decision Bench (~15.9k Korean items) and public Decision Index vs Decision-1, Jev, pplx-decider, Solar Decide, etc. (self-reported);
• Released under the commercial-friendly Kanana Open License.

Problem

Many agent steps are really "pick a tool / is this true / score it". Parsing prose from a general LLM is slow and brittle. Decision models productize that hop.

Peers

Same lane as TypeSafe AI's Jev, Microsoft Decision-1, Perplexity pplx-decider and Upstage Solar Decide. Kakao's pitch: multimodal, Korean in-house bench, commercial open license.

Limits

• Latency and board numbers are vendor-reported — re-eval cross-domain/language yourself;
• Does not replace long-form generative answers;
• ~3B class — hard reasoning may still need a larger model fallback.

Who it's for

• Teams doing agent tool-routing, ticket triage or product/image decisions;
• Anyone needing low-latency structured judgments with commercial OSS weights;
• Researchers following Kakao's Kanana line.

Getting started

1. Read https://www.kakaocorp.com/page/detail/12193 ;
2. Pull kanana-2-v-decide-3b-2611 from https://huggingface.co/collections/kakaocorp/kanana-2-v-decide ;
3. Wire probability / choice / score outputs into your router;
4. Calibrate on your label set and cross-check public Decision Index-style suites.