Krafton Releases Its In-House LLM 'Raon-1 Mini' for Free

Krafton released its in-house large language model (LLM), 'Raon-1 mini,' free of charge on the 21st. An LLM is an AI technology that understands and generates sentences like a human, and the newly released preview version is available for anyone to download and try for free. Raon-1 mini preview has been published on Hugging Face under the Apache 2.0 license, allowing it to be freely used not only for research purposes but also for commercial services and product development.

Raon-1 mini is a 30B-class model with 30 billion total parameters. It adopts a Mixture of Experts (MoE) architecture designed to activate only 4 billion of its total parameters during inference, combined with the company's proprietary linear attention technology. Krafton explained that this structure speeds up response times and reduces inference costs.

The model supports both Korean and English, and Krafton says it has achieved superior performance compared to similarly sized foreign models. The company cited natural Korean conversation, instruction-following capabilities such as formatting and persona settings, and agent applications for multi-step tasks as its key strengths. Detailed performance benchmarks will be released in a technical report when the official version launches.

"The release of Raon-1 mini demonstrates that we have independently developed a competitive language model built on more than five years of accumulated AI R&D capabilities," said Kang-wook Lee, Chief AI Officer (CAIO) of Krafton. "With a model that delivers higher performance at lower cost, we will expand AI adoption across a wide range of industries."

Krafton launched its AI model brand 'Raon' in April and has since released speech and vision models as open source. The company plans to continue improving Raon-1 mini toward its official release, and to publish a technical report detailing the training process and performance.