KRAFTON Drops Raon-1 Mini Preview

The compelling story is not the 29B headline; it is KRAFTON using architectural specialization to make long-context, reasoning-capable inference more practical locally. The preview looks promising, but developers should validate it against real workloads before treating it as a proven Qwen or Gemma replacement.

• –Six routed experts plus one shared expert activate per token, reducing compute while preserving a larger total capacity.
• –Eighteen of 25 layers use Raon Delta Attention, potentially reducing KV-cache pressure for long-context workloads.
• –NVFP4 support enables deployment on a single 32GB RTX 5090, while integrations cover Transformers, vLLM, llama.cpp, MLX, and agent tools. [Hugging Face model card](https://huggingface.co/KRAFTON/Raon-1-mini-preview)
• –KRAFTON’s benchmark table shows strong reasoning, coding, and Korean-language results, but performance is mixed across agentic and long-context evaluations.
• –Apache 2.0 covers the released model components; training data, methods, hosted API, and playground remain under separate terms. [Aju Press](https://www.ajupress.com/view/20261021170470164)