KRAFTON AI releases Raon-1 mini preview, an open 30B mixture-of-experts model

KRAFTON AI announced a preview of Raon-1 mini on October 20, a mixture-of-experts language model trained from scratch with about 30B parameters in total and 4B active. The weights are on Hugging Face as KRAFTON/Raon-1-mini-preview under Apache 2.0.

KRAFTON says its architecture matches Muse Glimmer on coding and agentic tasks while being faster and more memory-efficient. It reports a SWE-bench Verified score of 70.8 and a context of about 262K tokens; these are the lab's claims, not independent measurements. The post had about 540 likes on X by October 26.

Support in mainline llama.cpp is not there yet: a pull request was closed with a request that the vendor submit it, so local users run a forked build. One user's test on October 25 (IQ4_XS quantization, RTX 3080 10 GB) measured 1,417 tokens/s prompt processing and 45 tokens/s generation, and scored it 10/10 on a light-agent test and 30/30 on tool use, 77.6 on his own 48K "PA Core" suite against 60.3 for Qwen 3.6 35B A3B, and 57.3 on a change-of-plan task. These are one user's own benchmarks.