PUBG maker Krafton open-sources a 30B model, with its bug-fixing capabilities approaching Meta’s, and it leverages technologies from Kimi and DeepSeek at the base layer.

Beating AI News Brief: South Korean gaming giant KRAFTON has released a preview of Raon-1 mini, a Mixture of Experts (MoE) language model trained from scratch. The model has a total parameter count of around 28.8 billion, with only 4.3 billion activated per inference. It is optimized for Korean dialogue, programming, and agent capabilities, supporting a 260,000-token context window. Though built from scratch, its underlying technology draws on innovations from Chinese tech firms: it uses a hybrid attention architecture, with 18 of its 25 layers employing a linear attention mechanism adapted from Kimi Delta Attention. The team also simplified DeepSeek’s mHC hyperconnection technology to cut training computational costs. KRAFTON’s test results show Raon-1 mini scored 70.8% in the SWE-Verified code repair test, outperforming Qwen3.6-35B-A3B’s 63.2% and nearly matching Meta Muse Glimmer’s 72.0%. However, it lagged in the Terminal-Bench Lite terminal operation test, scoring 47.0%—well below Qwen3.6-35B-A3B’s 63.3% and Muse Glimmer’s 70.7%. For instruction following test IFBench, it hit 84.2%, higher than both comparison models. The model weights are open-sourced under the Apache 2.0 license, with versions available in GGUF, MLX, and other formats, supporting local run on Mac. The NVFP4 quantized version can be deployed on a single RTX 5090. The team also offers the DFlash2 speculative decoding model, which in tests on a single NVIDIA B300 delivers up to 2.18x higher output throughput than standard decoding.