KRAFTON Releases Raon-1-mini: A 30B MoE Bilingual Language Model
KRAFTON AI has released Raon-1-mini, a 30B-A4B Mixture-of-Experts language model trained completely from scratch under the Apache 2.0 license. The model features built-in reasoning, advanced tool-calling capabilities, and strong bilingual support for both Korean and English. This release provides the open-source community with a highly capable bilingual model optimized for both Korean and English tasks, agentic workflows, and math/science benchmarks. Its optimization features make advanced language models much more accessible for local deployment on consumer-grade hardware. Raon-1-mini supports speculative decoding via an included DFlash2 draft model and offers MXFP8 and NVFP4 quantization options. Notably, the NVFP4 format enables efficient inference directly on a single NVIDIA RTX 5090 GPU with 32 GB of VRAM.
## BACKGROUND
A Mixture-of-Experts (MoE) is a neural network architecture where multiple specialized sub-models (experts) are used, but only a subset of them is activated for each token to maintain high performance without high computational costs. Speculative decoding is an inference optimization technique that uses a smaller draft model to generate candidate tokens quickly, which are then verified simultaneously by the larger target model to reduce latency. NVFP4 is a low-precision 4-bit floating-point quantization format designed for NVIDIA Blackwell architecture GPUs to reduce memory bandwidth and storage requirements while preserving accuracy.