Oriel — AI & technology worth noticing, every day.

KRAFTON/Raon-1-mini-preview

AI model · https://huggingface.co/KRAFTON/Raon-1-mini-preview

First seen by Oriel on .

AI summary

KRAFTON's Raon-1 mini preview is an Apache 2.0, 30B-A4B MoE language model trained from scratch, targeting Korean conversation, reasoning, and agentic tasks.

  • Its hybrid architecture combines 18 Raon Delta Attention layers and 7 Gated MLA layers, supports 262,144-token context, and activates 4.3B parameters per token.
  • The model supports Korean, English, Japanese, Chinese, Spanish, French, and Thai, and tool calling with Hermes Agent, Pi, and OpenCode.
  • Benchmark results show strengths in math, science, instruction following, search, and Korean conversation, alongside weaker results in some coding, automation, and long-context comparisons.
  • A companion DFlash2 draft model reaches up to 2.18× output throughput over autoregressive decoding in tests on a single NVIDIA B300; results vary by hardware and workload.
  • Inference options include Transformers, vLLM, llama.cpp, and MLX; an NVFP4 checkpoint is documented for running on a single NVIDIA RTX 5090 with 32 GB VRAM.

#mixture of experts #language models #speculative decoding #llm agents

Generated by gpt-5.6-sol. Check the original for details.

Metric trends

Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained

Likes

Total likes reported by the source.

  • Hugging Face Models: 223, latest observation ; +109 over observed period since

Downloads (30d)

Rolling 30-day downloads. A change in this value is not new downloads during the observed period.

  • Hugging Face Models: 2,804, latest observation ; +2,779 over observed period since

Aliases

SourceExternal IDURL
Hugging Face ModelsKRAFTON/Raon-1-mini-previewhttps://huggingface.co/KRAFTON/Raon-1-mini-preview Read discussion