Kakao Open-Sources Kanana-2-v-decide-3b: 3.67B Multimodal Decision Model Returning Probabilities, Choices and Scores; Vendor Decision Index Avg 61.23

Core Background & Industry Pain Points

Agents repeatedly choose tools, classify tickets, or score options; generative LLMs emit text that must be parsed into JSON — slow and brittle. On 2026-11-06 Kakao open-sourced multimodal decision model Kanana-2-v-decide on Hugging Face for typed judgments under the commercial Kanana Open License.

Architecture Highlights & Internals

Built on Kanana-2-v-3b (~0.49B vision encoder + ~3.18B LM, ~3.67B total). State may be text, JSON, or chat history with up to 10 images and ~20,480 tokens. Questions and candidates are embedded into 2,560-d vectors; softmax over cosine similarities yields choice, noul, or score distributions. Candidate embeddings are cached across calls. Official paths: transformers system_one, decision pipeline, and a stdlib HTTP System One server.

Authoritative Benchmarks & Measured Scores

Vendor and model-card figures (not AICoder re-runs): 72.50 on Kakao Korean Kanana Decision Bench (15,909 questions); 61.34 on public Decision Index and 61.12 on Decision Index Vision (average 61.23), ahead of StartLux, Liquid AI d1 and Joey-1. Mean latency under 0.04s on one H200; Kakao claims about 43x faster than GPT-6 Luna. Press also cites Microsoft Decision-1 and TypeSafe Jev.

Developer Hands-on Guide

Load kakaocorp/kanana-2-v-decide-3b-2611 with trust_remote_code=True and call system_one for batched typed questions; images accept paths, URLs, base64 or PIL. Self-host with server_kanana_v_decide.py. See the HF collection. Caveats: scores and the 43x latency claim are vendor-reported; no video input; probabilities are relative to the supplied candidates.