Krafton Opens Raon-1 Mini Preview Under Apache 2.0

Krafton, the South Korean games company behind PUBG: Battlegrounds, has put its own large language model on Hugging Face. Raon-1 mini preview is a 30-billion-parameter mixture-of-experts model that its AI team says it designed and trained from scratch, released under the permissive Apache 2.0 licence, commercial use included.

What makes it notable is less the size than the ownership. GoTo’s Indonesian model this week, for instance, is built on Google’s Gemma and carries Google’s licence. Krafton built the architecture itself and is giving the weights away on standard open-source terms, with Korean conversation as its stated strength. The company is clear that this is an early cut: “This is a preview release, not a final model. Specs and results may change before the final version.”

Only 4.3bn parameters do the work on each token

The model holds 28.8bn parameters in total but activates 4.3bn per token, routing each one to six of 128 specialist “experts” plus a shared one. Most of its 25 layers use a linear-attention design Krafton calls Raon Delta Attention, which keeps a fixed-size memory state, and it handles up to 262,144 tokens of context. A compressed NVFP4 version runs on a single Nvidia RTX 5090 graphics card, according to the model card, and Krafton supplies a small draft model it says delivers up to 2.18 times the output speed on a single Nvidia B300, by its own measurements.

Krafton’s own benchmarks, scored with gpt-5.6-luna as judge, show a mixed picture against the two similar-sized open models it chose as rivals:

Benchmark	Raon-1 mini	Qwen3.6-35B-A3B	Gemma 4 26B-A4B
IFBench (instruction following)	84.2	66.6	80.9
SWE-Verified (coding)	70.8	63.2	61.6
MMLU Pro (knowledge)	76.5	85.9	82.9
Terminal-Bench Lite	47.0	63.3	52.0

Source: KRAFTON model card; its own evaluations.

Its headline Korean chatbot score comes from an internal test that has not been published.

The weights are open; the hosted service is not

Downloading requires a one-time request on Hugging Face, which Krafton says is approved automatically. The free playground and hosted API are a different matter: they sit under separate terms, not Apache 2.0, and those terms reserve Krafton’s right to charge fees later. Support in the popular llama.cpp runtime currently needs Krafton’s own branch, with upstream support pending.

The release builds on Raon, an AI brand Krafton launched in April with four open speech and vision models, Digital Today reported. Chief AI officer Kangwook Lee told the outlet the company had “independently developed a competitive language model” on more than five years of research. Aju Business Daily noted that Krafton already uses its AI work in games, including PUBG Ally, an AI teammate built with Nvidia that went into beta in Battlegrounds in June.

Krafton plans to keep updating the model before a full release, which it says will come with a technical report on training and final performance. That report will be the first chance to check the claims from outside.