KRAFTON AI, the artificial intelligence arm of the Korean game company KRAFTON (creator of PUBG), has released the preview of the Raon-1 mini model. It is an open-source LLM with a Mixture-of-Experts (MoE) architecture of about 30 billion total parameters, of which only about 4 billion are activated per token (30B-A4B).
What is Raon-1 mini?
The model was trained from scratch by KRAFTON AI and released under the Apache 2.0 license on Hugging Face. It has a special focus on natural Korean conversation, as well as good performance in English, programming tasks and agents. The architecture uses hybrid attention (linear and full) for long-context efficiency.
According to the company, Raon-1 mini is competitive with models like Meta Muse Glimmer 30B on coding and agent tasks, while being faster and more memory-efficient. There is a limited-time free playground and NVFP4 quantization support to run on a single RTX 5090.
Why it matters
This marks the entry of a major game publisher into the open models market, with emphasis on efficiency and Asian languages. The final version and larger models are planned for the future. Sources: official Raon site, IT Home and Chosun Biz reports.
Image credit: KRAFTON AI
By GeekikiBot
