AI Briefing
KOSign in

KRAFTON Releases Raon-1-mini-preview, a 30B-A4B MoE Model with Apache 2.0 License

·2026.10.21 16:54

Key point

The model supports speculative decoding via DFlash2 and offers NVFP4 quantization for 32 GB VRAM inference.

Details

  • KRAFTON AI released Raon-1-mini-preview, a 30B-A4B Mixture-of-Experts (MoE) language model trained from scratch and licensed under Apache 2.0.
  • The model demonstrates strong capabilities in Korean conversation and agentic tasks, with leading results on math and science benchmarks compared to other models.
  • It supports tool calling for agentic workflows compatible with Hermes Agent, Pi, and OpenCode.
  • Inference is accelerated by a provided DFlash2 draft model for speculative decoding.
  • Available in MXFP8 and NVFP4 quantization formats, with NVFP4 enabling operation on a single NVIDIA RTX 5090 GPU with 32 GB VRAM.
open-sourcemoekorean-llmkraftonraon-1-mininvidia-rtx-5090
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

RedditAgnes AI Releases Agnes-3.0-Qwen: 180B MoE Model with 1M Context and Strong Agentic BenchmarksRedditinclusionAI releases Ming-3.0-audio-flash, a 126B-parameter native audio-language modelKakaoKakao Releases Kanana-2-vision-encoder, Open-Sourcing Medium and Tiny WeightsGeekNewsArtCraft releases open-source IDE for AI image and video creation with 62-model catalogDeepSeekDeepSeek Releases V4.1-Flash-1017 and Opens Search & Fetch APIsPyTorchKRStrata Engine Released to Run Qwen3.8-Flash-Next 125B on 12GB VRAM