Raon-1 mini (GGUF)

Raon-1 mini (GGUF)

Runs a 28.8B mixture-of-experts model tuned for Korean chat, reasoning and tool-using agents in llama.cpp. GGUF copy of Raon-1 mini preview for people who run models on their own machine.

Made by KRAFTON on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This repository holds GGUF files of Raon-1 mini preview, a language model KRAFTON AI trained from scratch, so you can run it in llama.cpp. The files were converted from the BF16 weights. Q6_K and the smaller sizes were quantized with an importance matrix.

Who it suits: People who run language models locally and want one with strong Korean ability and tool calling. The authors call it a preview, so expect specs and results to change before a final version.

Industry | Who uses it | How they use it | Example
Software | Developers running local agents | Pair the model with an agent tool the card lists, such as OpenCode | A coding helper that works without a cloud service
Korean-language services | Teams working in Korean | Chat and question answering in Korean | A Korean assistant for internal questions
Research | Model researchers | Try a mixture-of-experts model that uses few parameters per token | Comparing it with a dense model of similar size
Hobby use | Enthusiasts | Chat with a large model on their own computer | An offline chat in Korean or English

You can skip it if you need a final, stable release, or if you cannot spare tens of gigabytes of disk and memory.

Where to use it:

• llama.cpp on Windows, macOS or Linux
• Agent tools the model card names, Hermes Agent, Pi and OpenCode
• Your own computer, with no hosted service needed to run the files

Use case | Who | What they get back
Korean conversation | Korean-language teams | Replies in Korean, with English also supported
Step-by-step reasoning | Developers | Answers where the model thinks before it replies
Agent workflows | Developers | Tool calls an agent tool can act on

How to set it up:

• Python, to install the Hugging Face client
• llama.cpp, which loads the GGUF files
• Disk space, about 14.4 GB at 4-bit by Hugging Face's estimate

1. Install the Hugging Face client.
2. Download the repository files to your computer.
3. Point llama.cpp at the GGUF file whose quantization size fits your memory.

pip install -U huggingface_hub
hf download KRAFTON/Raon-1-mini-preview-gguf

How to use it: Open a GGUF file in llama.cpp and chat with it in Korean or English. The model thinks before it answers and can call tools for agent workflows. The source gives no llama.cpp command, so follow llama.cpp's own guide for loading a file.

Tool | What it does
llama.cpp | Loads the GGUF files and runs the model locally
hf download | Fetches the repository files from Hugging Face
Q6_K and smaller | Quantized sizes made with an importance matrix

Example: You ask the model a question in Korean and it works through the problem before giving its answer. You can also connect it to an agent tool such as OpenCode so it calls tools on your behalf.

For newcomers and experts: If you are new to local models, download the files and start with the smallest quantization that fits your memory. A smaller file runs on more machines and loses some quality.

If you know the field: the model has 25 layers, 18 of them Raon Delta Attention with a fixed-size state and 7 gated MLA layers that keep a KV cache. It routes each token to 6 of 128 experts plus 1 shared expert, so 4.3B of 28.8B parameters are active. Context reaches 262,144 tokens.

Hugging Face repository
GGUF file
llama.cpp
Local chat or agent

Good to know: Only a fraction of the parameters work on each token, which is why a model this large can be practical to run.

Title: Parameters in billions
Total: 28.8 B
Active per token: 4.3 B

• Preview: this is an early release, so details may shift once a final model ships
• Context: up to 262,144 tokens
• License: Apache 2.0
• Cost to run: none named, though the card mentions a free but time-limited hosted playground
• Data it reaches: the files you download to your computer, nothing else named
• Not for: a final, stable release, since this is a preview

On Hugging Face, 13 November 2026

• Task: text generation
• Parameters: 28.8B
• Weights: about 14.4 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 44,362
• Likes: 77
• License: apache-2.0
• Based on: KRAFTON/Raon-1-mini-preview
• Tags: gguf, raon, llama.cpp, text-generation, imatrix, conversational

Links

• Model: https://huggingface.co/KRAFTON/Raon-1-mini-preview-gguf
• Files: https://huggingface.co/KRAFTON/Raon-1-mini-preview-gguf/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download KRAFTON/Raon-1-mini-preview-gguf

Then follow the usage in its model card. Nothing in this kit runs by itself.