Krafton Releases Preview of 'Raon-1 mini', a Self-Developed 30B-Class High-Efficiency LLM

What happened

On October 21, Krafton released a preview version of Raon-1 mini. It is a large language model (LLM), the kind of AI that reads and writes text. Krafton built the model from scratch, designing it and training it on its own. The model is open source under the Apache 2.0 license, so anyone can use it for research, commercial services, or products. It is shared on Hugging Face, a popular website for AI models. It supports English and Korean.

Key numbers and facts

• Total size: 30 billion parameters (30B). Parameters are the numbers inside an AI model that store what it has learned.
• Active size: about 4 billion parameters (4B) are used at one time, out of the 30B total.
• Design: a mixture-of-experts (MoE) structure, combined with Krafton's own linear attention technology.
• Releases: 4 AI models were released as open source in April under the Raon brand.
• Research: 54 papers published at major AI conferences so far. 9 main-track papers were accepted at NeurIPS 2026 this year.

How the model is built

Most big AI models use all of their parameters for every answer, which costs a lot of computing power. Raon-1 mini uses a mixture-of-experts design. This means it has many smaller 'expert' parts, but only a few work on each question. Krafton says this makes answers faster and cheaper to run. Its performance is strong for its size, and on some tests it gets close to large Korean models.

Why it matters

Krafton says the real test of an AI model is not just its test scores, but how much good work it does for the money spent to run it. Krafton focuses on Korean conversation, AI agents that can follow multi-step instructions, and fast replies. Because the model is free to use, companies and developers can build their own AI tools on it at lower cost.

What to watch

• The full, official version of Raon-1 mini, which Krafton plans to release after more updates.
• A technical report with details on how the model was trained and how well it performs, to be released with the official version.
• How developers use the open-source model in real products.