Kanana-2-v-decide brings local AI decisions, with license limits
Kakao released Kanana-2-v-decide on November 6, giving developers downloadable model weights for repeated decisions involving text and images. The release targets tasks such as sorting customer requests and selecting an agent’s next tool. Kakao plans to use it in internal search and product classification projects, according to its announcement.
The practical opportunity is to run a focused decision model inside an application. The commercial boundary deserves an early check: Kakao’s custom license permits many own-service uses, while offering the model to third parties through several distribution routes requires separate permission.
What Kanana-2-v-decide returns
The model card describes a 3.67-billion-parameter system that accepts text, structured data or chat history, with up to 10 images and 20,480 tokens. It produces choices, yes-or-no probabilities or ordered scores. It does not generate prose or accept video.
A useful distinction is that its probabilities depend on the supplied candidates. Adding or changing an option changes the distribution. A high number therefore needs to be tested against the application’s actual categories and mistakes before it becomes an automation threshold.
For example, a retailer could use a product photo and description to suggest a catalog category, then send uncertain suggestions to an editor. That is an illustrative workflow, not a test performed by ByteForward. The application still needs rules for missing information and cases where none of its categories fit.
Where commercial rights stop
The Kanana Open License draws a specific distinction. Section 4.2 allows development and operation of your own services without a separate commercial license when Section 4.1 does not apply. Section 4.1 requires Kakao’s separate grant for offering or reselling model access through API, cloud or other remote services, system-integration or on-premise solutions, and embedded on-device distribution.
Redistribution also carries conditions, including the license and notice, marking modified files, passing on use restrictions and displaying the required attribution. Distributed AI models trained or improved using the materials or outputs must include Kanana as a name prefix. These are open weights under a custom license. Teams planning a product should check their exact distribution arrangement against the full terms.
What a local deployment needs
The published reference environment specifies Python 3.12, CUDA, BF16 and SDPA, with PyTorch 2.10.0 and Transformers 5.5.3. The weights file is 7.33 GB. File size is not a complete GPU-memory budget: runtime allocations and inputs also need room.
The supplied HTTP server exposes the System One API and serializes model inference, running one model call at a time. Its default bind address is localhost. Authentication is configurable through DECISION_API_KEY, and the documentation warns to set a key before making the service reachable from other computers. This reference server needs deployment planning before it serves concurrent application traffic.
Input handling also matters. The image processor converts images to RGB, composites transparency onto white and resizes images before patching them. A pilot should include the actual product photos the service will receive, including small details and transparent backgrounds.
How to read Kakao’s speed and benchmark claims
Kakao’s technical account reports about 42 milliseconds for one decision on Korean product-review data using one H200 GPU. Token count affects latency. Its claimed 43-fold advantage over GPT-6 Luna comes from that comparison, not independent testing by ByteForward or a guarantee for other hardware.
Kakao explains that it embeds questions with their input and compares them with answer embeddings. Reusing candidate embeddings can reduce repeated work when an application keeps the same categories. That is a reason to test warm and cold requests separately, rather than copying a single latency number into a capacity plan.
The model card reports a 61.23 average across its two public decision benchmarks. In that table, Joey-1 scores higher on the text benchmark alone. The technical blog presents a different comparison set and slightly different Kanana scores. The two tables should not be merged into one leaderboard or treated as proof of universal superiority.
Start with the decisions your application makes
A useful first evaluation would keep a held-out set of labeled examples, measure category mistakes, and record latency under expected traffic. Test whether uncertain cases reach the intended review path. Repeat the evaluation when category descriptions change, because changing the options changes the problem being scored.
Teams comparing self-hosting with a managed route can also read our coverage of Liquid AI’s d1 Ultra decision model. Its hosted billing model is a separate tradeoff. For Kanana, the immediate decisions are whether the license fits the product and whether the model makes reliable choices on the team’s own data.
Featured image: AI-generated editorial illustration by ByteForward. Kanana logo: Kakao. Conceptual depiction, not a product interface.
