Kakao's Multimodal AI Model 'Kanana-v' Reaches 1.6 Million Downloads on Hugging Face

Kakao's multimodal artificial intelligence (AI) language model, 'Kanana-v,' has recorded 1.6 million downloads on the global open-source AI platform Hugging Face.

According to Kakao on July 9, the open-source vision-language model (VLM) 'Kanana-1.5-v-3b-Instruct' was downloaded 380,000 times on Hugging Face in the past month alone. The cumulative number of downloads for this model has now reached approximately 1.6 million.

The 'Kanana-1.5-v-3b-Instruct' model released on the open-source artificial intelligence (AI) platform Hugging Face. Kakao

VLM refers to AI models that can process and understand both text and images. They can be used for document analysis, image description, and interpretation of visual information.

In terms of downloads over the past month, Kanana-v outperformed its competitors. It was downloaded 2.8 times more than Naver Cloud's similarly sized VLM 'HyperCLOVA X-Seed-Vision-Instruct-3B' (about 137,000 downloads), and also surpassed the larger 'Exaone-4.5-33B' from LG AI Research (about 247,000 downloads).

Kanana-v is officially listed as a supported model on the open-source inference framework 'vLLM,' allowing immediate deployment in services without separate configuration. As a lightweight model, it can be operated economically even in environments with limited GPU resources.

Kakao has been focusing on enhancing multimodal understanding, instruction execution, and reasoning capabilities of its AI models. After releasing 'Kanana-1.5-v-3b' in July last year, the company introduced 'Kanana-v-4b-Hybrid' in January this year, which minimizes hallucination through self-checking. Currently, after optimizing 'Kanana-2' for agentic AI, Kakao is developing 'Kanana-2.5' and working on tokenizer development to further improve inference speed and cost efficiency.

A Kakao representative stated, "Multimodal models play a key role in the completeness of agentic AI, which fully understands user context and autonomously completes complex tasks," adding, "We plan to continue our innovation so that more developers and companies can implement high-performance, high-efficiency agentic AI services without cost burden."

This content was produced with the assistance of AI translation services.