Krafton is expanding its AI capabilities by open-sourcing its self-developed large language model (LLM). The company aims to create high-efficiency AI models applicable across various industries.
Krafton unveiled a preview version of its Raon-1 Mini, a large language model with 30 billion parameters, on October 21.
The Raon-1 Mini preview is an LLM that Krafton developed entirely in-house, from model design to training from scratch. It supports both English and Korean.
Krafton designed a proprietary model architecture that activates only 4 billion of the total 30 billion parameters, utilizing a mixture of experts (MoE) structure combined with its own linear attention technology. This approach achieves fast response times and low inference costs while ensuring high performance and efficiency.
In April, Krafton launched its AI model brand, Raon, and previously open-sourced four types of AI models, including a speech language model and a real-time voice conversation model. The Raon-1 Mini preview serves as the foundation for future AI models developed by Krafton, which has released it under the Apache 2.0 license on Hugging Face.
Krafton stated that this development has equipped them with the technical capability to design performance and inference efficiency tailored to various AI application needs.
The Raon-1 Mini preview has demonstrated superior performance compared to similar international models, approaching the capabilities of large domestic models in some metrics. It particularly excels in natural Korean dialogue, emphasizing its ability to accurately follow instructions such as formatting and persona settings. This instruction-following capability is essential for agents that can handle multi-step tasks independently.
Krafton highlighted cost efficiency as a competitive advantage of this LLM. A company representative noted, "As AI utilization spreads across industries, we believe that performance and cost-effectiveness in real-world applications determine a model's competitiveness. Therefore, we developed the model with practical usability in mind rather than focusing solely on benchmark scores."
Krafton plans to continuously update the Raon-1 Mini to enhance its performance before releasing a final version, which will include a technical report detailing the model's training process and final performance.
Lee Kang-wook, Krafton's Chief AI Officer, stated, "The release of Raon-1 Mini is a result of Krafton's over five years of accumulated AI research and development capabilities, allowing us to independently develop a competitive language model. We aim to enhance performance while reducing cost burdens, thereby broadening the scope of AI utilization across various industries."
Krafton is also applying its AI research achievements to gaming. In June, it beta-tested an AI companion called 'PUBG Ally,' developed in collaboration with NVIDIA, in its game Battlegrounds, allowing users to interact with AI characters for item collection and strategy planning.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.