PUBG developer open-sources 30B model: bug fixing approaches Meta, built on Kimi and DeepSeek technology

Beating AI News Flash: South Korean gaming giant KRAFTON releases the Raon-1 mini preview. This is a MoE language model trained from scratch, with approximately 28.8 billion total parameters, activating only 4.3 billion per inference, with a focus on optimizing Korean dialogue, coding, and Agent capabilities, supporting a 260K token context.

Although trained from scratch, the underlying technology also bears the imprint of Chinese vendors. It adopts a hybrid attention architecture, with 18 of its 25 network layers using a linear attention mechanism improved from Kimi Delta Attention. Additionally, the team simplified DeepSeek's mHC hyper-connections technology to reduce training compute overhead.

Tests published by KRAFTON show that Raon-1 mini scored 70.8% on the SWE-Verified code repair test, surpassing Qwen3.6-35B-A3B's 63.2% and approaching Meta Muse Glimmer's 72.0%. However, on the Terminal-Bench Lite terminal operation test, it scored only 47.0%, significantly trailing Qwen3.6-35B-A3B's 63.3% and Muse Glimmer's 70.7%. On the IFBench instruction-following test, it reached 84.2%, higher than the above two models.

The model weights have been released under the Apache 2.0 license, with GGUF, MLX, and other versions available, supporting local operation on Mac. The NVFP4 quantized version can be deployed on a single RTX 5090. The team also provides the DFlash2 speculative decoding model, which in tests on a single NVIDIA B300 achieved up to 2.18x higher output throughput compared to normal decoding.