AI BriefingKakao Releases Open-Source 3B Multimodal Decision-Making Model 'Kanana-2-v-decide'
·2026.11.05 08:00
Key point
Kakao has released the 3B-scale multimodal decision-making model Kanana-2-v-decide as open source.
Details
- Kakao has released Kanana-2-v-decide, a 3B-scale multimodal decision-making model that understands text and images, as open source
- It recorded a response time of approximately 42ms on a single H200 GPU, achieving a speed 43 times faster than GPT-6 Luna
- It scored 72.62 on the in-house evaluation set (Kanana Decision Bench), outperforming Microsoft-Decision-1 (71.98) and Jev 1.14.0 (71.66)
- It applies an embedding-based structure, allowing expansion without limiting the number of options, and reduces computational costs for repetitive tasks through vector reuse
Technical Features and Performance
- It returns judgment results in the form of Probability (Noul), Choice, and Score, enabling immediate integration with service logic
- Through MOPD (Multi-teacher On-Policy Distillation) training, it reduces unnecessary loss choices in safety evaluations by up to 73%
- It improves reliability by lowering the Expected Calibration Error (ECE) from 2.06% to 1.12% through Confidence calibration
Application and Release
- It supports APIs compatible with Jev and can be called directly via typesafe-sdk
- It has released model checkpoints on Hugging Face to support direct execution and fine-tuning by researchers and enterprises
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
spinf-serve: Open-source inference server for bulk LLM scoring on single GPU