KRAFTON releases Raon-1 mini preview, a 30B-A4B MoE model
Original title: KRAFTON/Raon-1-mini-preview — new model trending #30 on Hugging Face
The Apache 2.0 preview activates 4.3B of 28.8B parameters and supports a 262,144-token context.
KRAFTON AI released Raon-1 mini preview, an open Mixture-of-Experts language model trained from scratch and licensed under Apache 2.0. It targets Korean conversation and agentic work and was trained on Korean, English, Japanese, Chinese, Spanish, French, and Thai. KRAFTON says the card is a preview, so specifications and benchmark results may change. NVFP4 weights are offered for inference on one NVIDIA RTX 5090 with 32 GB of VRAM.
Unconfirmed: KRAFTON says this preview is not final and that specs and results may change.
- Six of 128 routed experts plus one shared expert activate per token.
- Eighteen of 25 layers use Raon Delta Attention; seven use gated MLA.
- A DFlash2 draft model is included for speculative decoding.
- MXFP8 and NVFP4 quantizations are available.
- Tool calling supports Hermes Agent, Pi, and OpenCode.
- Vendors
- KRAFTON
- AI models
- Raon-1 miniDFlash2
From the original article
5 min read at huggingface.cotext-generation · 917 downloads · 154 likes · license apache-2. 0
By KRAFTON · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.