Kakao Releases Open-Source Lightweight Multimodal AI 'Kanana-1.5-v-3b' Available for Commercial Use

Kakao has released 'Kanana-1.5-v-3b', a lightweight multimodal language model licensed under the commercially usable Kanana License. This model features approximately 3.6 billion (3.6B) parameters and utilizes a ViT-based Vision Encoder alongside its proprietary C-Abstractor architecture. It demonstrated high performance compared to competing models on Korean-specific benchmarks (such as KoOCRBench), particularly scoring 85.93 on KoOCRBench, significantly surpassing the second-place Qwen2.5-VL (50.67). With an average score of 74.00 on English benchmarks, it has secured global competitiveness. The training process employed knowledge distillation (KD) techniques to transfer knowledge from a larger model (Kanana-1.5-v-9.8b), along with instruction following and Direct Preference Optimization (DPO) using high-quality Korean datasets (KoMIF, KoText-HQ). The model is available for download on Hugging Face, and Kakao is currently developing future models with enhanced reasoning capabilities and an integrated multimodal model, 'Kanana-o'.