KRAFTON, the company behind PUBG, releases Raon-1 mini, an open model built from scratch
IA en un minuto newsroom · Editor: Jon Elgezabal
In 30 seconds
Raon-1 mini, the language model KRAFTON has built without starting from another one, can be downloaded from Hugging Face and used even in commercial products. Its mixture of experts only puts a small share of the parameters to work at each step, and the NVFP4 variant fits on a single RTX 5090. The company warns that this is a preview and that everything may change before the final release.
KRAFTON, the South Korean video game company behind PUBG, has released on Hugging Face (the platform where AI models are shared) the preliminary version of Raon-1 mini, a language model it has designed and trained from scratch. It has open weights under the Apache 2.0 license, so anyone can download and use it, for commercial purposes too, after requesting access on its page, which is granted automatically. The license does not cover the company's hosted API or its test playground, which are in a free private beta and are requested by email.
It has 28.8 billion parameters, but with its mixture-of-experts architecture (MoE, which divides the work among specialized parts of the model) it only uses 4.3 billion for each token: of its 128 experts, 6 are activated, plus a shared one. According to the company's statement as reported by Newspim, this gives it faster answers and a lower cost. It handles up to 262,144 tokens of context (the amount of text it can keep in mind at once) and comes in two compressed formats, MXFP8 and NVFP4; with the second one it runs on a single NVIDIA RTX 5090 graphics card, with 32 GB of memory.
KRAFTON highlights its conversation in Korean and agent tasks, in which the AI chains steps together and uses tools on its own. It was trained in Korean, English, Japanese, Chinese, Spanish, French and Thai, and it works with the Hermes Agent, Pi and OpenCode agents through an API compatible with OpenAI's. In the tests the company publishes, it beats Qwen3.6-35B-A3B and Gemma 4 26B-A4B, two models of similar size, at following instructions (IFBench: 84.2, against 66.6 and 80.9) and at search (BrowseComp Plus: 55.7, against 34.8 and 32.2). It trails in others, such as Terminal-Bench Lite (47.0, against 63.3 and 52.0) or MMLU Pro (76.5, against 85.9 and 82.9). In Korean conversation it scores 92.9, but on an internal test the company has not made public.
It is a preview, not the final model: the company warns that the specifications and results may change, and it will publish a technical report on the training and the final results when it launches the definitive version. KRAFTON launched the Raon brand in April with four voice and image models, and Raon-1 mini will be the base for the next AI models it develops. The company already uses AI in its games: in June it brought PUBG Ally, an AI companion built with NVIDIA, to Battlegrounds in beta, according to Aju News.
Why it matters · analysis and opinion
A video game studio releasing a language model of its own, rather than an adaptation of someone else's, shows that training from scratch is no longer the preserve of the big labs. For a company, the practical part is the combination of a commercial license and modest hardware needs: it can be tested on its own server, without sending data outside and without paying per use. Its stated strength is Korean, so the immediate interest is greater for anyone working with that market; there are no published results in Spanish, and the sensible move is to measure it first with real cases. Being a preview, it suits trials and prototypes, and it is worth waiting for the technical report before building a product on it.
Official source: KRAFTON on Hugging Face · Written with the help of AI: how we make the news
