Oriel — AI & technology worth noticing, every day.

KRAFTON/Raon-1-nano

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

First seen by Oriel on .

AI summary

KRAFTON AI’s Raon-1 nano is an Apache 2.0 bilingual language model with 1.5B effective parameters, long-context support, tool calling, and local inference formats.

  • The dense model has 24 layers, combining 18 Raon Delta Attention layers with six Gated MLA layers and a fixed-size recurrent state.
  • It supports up to 262,144 tokens, while only the six full-attention layers retain a KV cache of approximately 7 KiB per BF16 token.
  • Raon-1 nano was derived from Raon-1 mini preview through MoE-to-Dense pruning, knowledge distillation, 3T-token continued pre-training, SFT, OPD, and RL.
  • It leads the listed comparison models on several instruction-following, context-following, Korean conversation, and Korean tool-use benchmarks, but not every benchmark.
  • Transformers, vLLM, llama.cpp, MLX, and LiteRT-LM enable deployment across NVIDIA, AMD, CPU, and Apple silicon environments, with LiteRT-LM limited to 4,096 tokens.

#language models #korean ai #tool calling #on-device inference

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: 145, latest observation ; +44 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: 1,510, latest observation ; 0 over observed period since

Aliases

SourceExternal IDURL
Hugging Face ModelsKRAFTON/Raon-1-nanohttps://huggingface.co/KRAFTON/Raon-1-nano Read discussion