Act on this, mostly signal

KRAFTON's Raon-1 mini: 30B MoE under Apache 2.0 runs on one RTX 5090

What to do

Try the NVFP4 quantization on a 32 GB RTX 5090 if you run local models.

Why

It is a 30B-A4B mixture-of-experts model with a DFlash draft model for speculative decoding.

What the research found

KRAFTON released Raon-1 mini preview, a 30B-A4B MoE model under Apache 2.0 with NVFP4 and MXFP8 quantizations, NVFP4 fitting on a single RTX 5090 32 GB, and a DFlash draft model for speculative decoding.

Claims and how well the evidence supports them

Raon-1 mini is a 30B-A4B mixture-of-experts model released under Apache 2.0 98%
NVFP4 quantization enables inference on a single RTX 5090 with 32 GB VRAM 98%
A DFlash draft model is provided for speculative decoding 78%

Evidence

reddit.com https://reddit.com/r/LocalLLaMA/comments/1xbh0ev/kraftonraon1minipreview_hugging_face
huggingface.co https://huggingface.co/KRAFTON/Raon-1-mini-preview-gguf

Actionability rated medium.