Is Kanana-2-v-decide state of the art?Pending benchmarks.Announced Nov 5, 2026; no independent results yet.
Kakao's open-weight multimodal decision model (checkpoint kanana-2-v-decide-3b-2611, 3.67B parameters: 0.49B vision encoder + 3.18B language model) built on Kanana-2-v-3b. Given a state (text, JSON or chat, optionally with images) it answers typed questions — choice, noul (yes/no) or score — and returns a probability for every option instead of generating text. Self-host via Transformers or the included System One HTTP server; no public hosted API is stated. Weights on Hugging Face under the Kanana Open License.
The details so far
- Lab
- Kakao
- Announced
- Nov 5, 2026
- API
- Not yet
- Open weights
- Published
- Parameters
- 3.67B (vision 0.49B + LM 3.18B)
- Role
- Decision model: probability distributions over choice / noul / score questions about a multimodal state; no text generation
- Checkpoint
- kanana-2-v-decide-3b-2611 (kakaocorp/kanana-2-v-decide-3b-2611 on Hugging Face)
- Base
- Built on Kanana-2-v-3b (Kanana-2-3b language model + Kanana vision encoder); separate from other Kanana releases
- Inputs
- Text, JSON or chat messages; up to 10 images; up to 20,480 tokens. Video not supported
- Embedding
- 2,560-dimensional embeddings; softmax over cosine similarities between question+state and candidate answer embeddings
- Serving
- Transformers AutoModel.system_one and a stdlib HTTP server (POST /v1/systemone) for self-host; binds 127.0.0.1 by default. No public hosted Kakao API stated
- License
- Kanana Open License (HF license tag: other / license_name kanana-open-license)
- Related
- Separate from Kanana-2-v-3b and other Kanana foundation / decide naming
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: Decision Index 0.4.1 public 61.34 and Decision Index Vision 0.4.1 public 61.12 (average of those two 61.23) on the model card comparison table.
Lab claim: Kanana Decision Bench (Kakao in-house Korean text set, 15,909 questions) score 72.50 (model card).
Lab claim: mean response under 0.04 s on one H200 GPU; Kakao press compares that latency to GPT-6 Luna as about 43× faster.
Sources
LABKakao press: Kanana-2-v-decide open-source releasekakaocorp.com/page/detail/12193 · Nov 6, 2026REPOkakaocorp/kanana-2-v-decide-3b-2611 model card (Hugging Face)huggingface.co/kakaocorp/kanana-2-v-decide-3b-2611REPOkakaocorp/kanana-2-v-decide collection on Hugging Facehuggingface.co/collections/kakaocorp/kanana-2-v-decideTracking continues until an independent board lists it.