Krafton releases 30B-class Raon-1 Mini LLM on Hugging Face
The preview model is available under the Apache 2.0 license for free download, research and commercial use, and supports Korean and English.
Krafton released a preview version of its self-developed 30-billion-parameter large language model, Raon-1 Mini, on the AI model platform Hugging Face on October 21, according to techm.kr. The model is published under the Apache 2.0 license, and users can download it free of charge.
The license covers research as well as commercial services and product development. Raon-1 Mini supports Korean and English.
The model combines a mixture-of-experts (MoE) architecture with Krafton's self-developed linear attention technology. Of its 30 billion total parameters, only 4 billion are activated during inference.
According to Krafton, the preview showed strong performance against overseas models of similar size and approached domestic large models on some metrics. Krafton said it plans to keep updating the preview before releasing a full version, alongside a technical report describing the training process and final performance.
Krafton CAIO Lee Kang-wook said the release reflects more than five years of accumulated AI research and development capability. The company has built up that R&D base since before this release, he said.
Quick answers
Is Raon-1 Mini free to use?
Yes. It is published on Hugging Face under the Apache 2.0 license, which allows free download and use for research, commercial services and product development.
What languages does Raon-1 Mini support?
Krafton's Raon-1 Mini supports Korean and English.
How large is Raon-1 Mini?
It has 30 billion total parameters, with 4 billion activated during inference under a mixture-of-experts architecture.