KRAFTON Pioneers in AI Development with Raon-1 mini: A Breakthrough in Foundation Models
In an exciting development within the artificial intelligence community, South Korean gaming giant KRAFTON—best known for creating PlayerUnknown’s Battlegrounds (PUBG) and owning the studio behind the acclaimed game Subnautica—has announced the release of a groundbreaking AI model: Raon-1 mini. This new foundation model, boasting 30 billion parameters, signifies a significant stride forward in large-scale AI research and development.
From Scratch: A Monumental Achievement
What sets Raon-1 mini apart is that KRAFTON claims to have trained it entirely from the ground up. Utilizing an original architecture that incorporates some of the latest advancements in AI research, the company has demonstrated impressive technical prowess. Training a competitive large language model (LLM) of this scale from scratch is an endeavor fraught with both technical challenges and immense financial costs, often limiting such initiatives to tech giants with vast resources like Google or OpenAI.
A Disruptive Entry in the AI Landscape
The significance of KRAFTON’s achievement becomes clearer when considering the context of current AI benchmarks. Compared to established models such as Qwen—a well-known MoE (Mixture of Experts) model—Raon-1 mini shows itself to be not only competitive but in several respects possibly superior. Early results indicate that Raon-1 mini can rival, and in some cases outperform, models like the similarly sized Qwen MoE and even challenge larger dense models such as Qwen3.8-27B.
While independent benchmarks are awaited to provide a definitive comparison, the initial data is promising and suggests that KRAFTON’s model is poised to introduce fresh competition in the open-model AI arena.
Rapid Deployment and Accessibility
In addition to the impressive technical achievement, KRAFTON has taken steps to make Raon-1 mini broadly accessible. The model has been released with support for multiple platforms, including NVFP4 and Apple Silicon MLX, along with weights licensed under the permissive Apache 2.0 license. This openness facilitates further research, development, and integration by the wider AI community.
For those interested in exploring Raon-1 mini firsthand, the model and benchmark data are available on Hugging Face:
Raon-1 mini Preview on Hugging Face
Implications for the Future
KRAFTON’s foray into advanced AI research signals an exciting shift where even game development companies with specialized expertise can meaningfully contribute to the evolution of foundation models. This could herald a more democratized and competitive landscape for AI development, accelerating innovation across industries.
As an AI enthusiast and developer, I am personally eager to begin optimizing and building upon this model to further explore its capabilities. The advent of Raon-1 mini hints at a future where open, high-performance foundation models become more accessible, fostering a new wave of innovation and competition.
Conclusion
KRAFTON’s successful development of Raon-1 mini underscores the growing democratization of AI research. By demonstrating that a gaming company can train a high-quality foundation model from scratch, this milestone challenges assumptions about resource limitations and opens new opportunities for the broader AI community. The coming months will be pivotal in observing how Raon-1 mini influences the competitive landscape and advances the field at large.
