KRAFTON releases Korean MoE model Raon-1 mini preview

KRAFTON AI released Raon-1 mini preview, a 30B-A4B MoE model trained from scratch under Apache 2.0. It targets Korean conversation, reasoning, instruction following and agentic tool calling, ships a DFlash2 speculative-decoding draft model, and its NVFP4 quantization runs on a single 32 GB RTX 5090.

• Eagle-4 Released: Faster than DFlash2 & DSpark2026-10-21
• China-modified RTX 5090 with 96GB Appears on Alibaba2026-09-12