KRAFTON AI Open-Sources Raon-1 Mini: A 30B MoE Model Rivaling Meta's Muse Glimmer
KRAFTON AI has released and open-sourced the preview of Raon-1 mini, a model trained from scratch. Featuring a 30B-A4B Mixture-of-Experts architecture with linear attention and speculative decoding, it is licensed under Apache 2.0. The company claims performance comparable to Meta's Muse Glimmer 30B in coding and agentic tasks, while offering faster inference and better memory efficiency.
Full Report
October 25 — KRAFTON AI, the artificial intelligence division of South Korean gaming giant KRAFTON, has released a preview version of its Raon-1 mini model, trained from scratch. The official release and larger subsequent models are expected soon.
Raon-1 mini features a Mixture-of-Experts (MoE) architecture, incorporates linear attention, and supports speculative decoding, with an overall size of 30B-A4B. While focusing on Korean language performance, it also maintains strong English capabilities. The model is open-sourced under the Apache 2.0 license. KRAFTON AI is also offering a limited-time free API trial on its official website.
According to KRAFTON AI, the preview version of Raon-1 mini matches Meta’s Muse Glimmer 30B in coding and agentic tasks, while offering faster speeds and higher memory efficiency. Additionally, it enables natural conversational interactions.
References