Raon-1 mini

Raon-1 mini

Chats and reasons, with a strength in Korean, and runs tool-calling agent tasks. A 28.8B mixture-of-experts model that activates 4.3B parameters per token with a 262,144-token context. A preview release from KRAFTON AI.

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: Raon-1 mini is a mixture-of-experts language model that KRAFTON AI trained from scratch. It has 28.8B parameters in total and uses 4.3B for each token. The authors point to Korean conversation and agent tasks as its strengths. The release is a preview, which means the authors may adjust the numbers before a final model appears.

Who it suits: It suits developers who want an open model with good Korean and tool calling, and who have a GPU that fits its compressed formats.

Industry | Who uses it | How they use it | Example
Korean-language products | App developers | Build chat and assistant features that work well in Korean | Answering customer questions in Korean
Agent building | Agent developers | Run tool-calling workflows through agent tools such as OpenCode | Letting an agent call a tool during a task
Local AI hobby | Enthusiasts | Run a mid-sized model on one consumer GPU in NVFP4 | Chatting on a single RTX 5090
Research | Model researchers | Study its hybrid linear and full attention design | Testing long inputs up to 262,144 tokens

You can skip it if you need a final, stable release, since the authors call this one a preview.

Where to use it:

• Python with the transformers library
• Agent tools that the card says it supports, namely Hermes Agent, Pi and OpenCode
• A single NVIDIA RTX 5090 with 32 GB of VRAM, using the NVFP4 format
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Korean chat | App builders | Conversational answers in Korean, English and other trained languages
Agent tool calls | Agent developers | Tool-calling output for agentic workflows
Faster generation | Engineers | Speculative decoding with the provided DFlash2 draft model

How to set it up: Plan on Python, the download tool, and a GPU with space for the weights, listed at 28.8 GB in an 8-bit float format.

• Python on Windows, macOS or Linux
• The huggingface_hub package
• The transformers library to run the model
• A GPU with enough memory, or the NVFP4 format for a 32 GB card

1. Install huggingface_hub.
2. Download the model files to your computer.
3. Follow the model card's run section to load it.

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

How to use it: The visible card gives the download step and no run command. Once loaded, you chat with the model, which has built-in thinking, or connect it to an agent tool for tool calling. A free hosted playground was offered for a limited time.

hf download KRAFTON/Raon-1-mini-preview

Example: You ask the model a question in Korean. It thinks first with its built-in reasoning, then replies in Korean. In an agent tool it can also call a tool and use the result.

For newcomers and experts: If you are new to this, it is a large model that only wakes part of itself for each word, which keeps it fast. A first small try is the hosted playground the card mentions, if it is still open.

For people who know the field, 18 of its 25 layers use Raon Delta Attention, which keeps a fixed-size state. The 7 full-attention layers keep a KV cache of about 8 KiB per token in BF16. Each token uses 6 of 128 routed experts plus one shared expert.

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

Your prompt goes in
A router picks 6 routed experts and 1 shared expert
The hybrid attention layers process the context
The model thinks, then answers or calls a tool

Good to know: Seven languages went into training: Thai, French, Spanish, Chinese, Japanese, English and Korean, with Korean the strongest.

• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the prompts you send and the downloaded model files
• Not for: use as a final model, since the card says specs and results may change

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 28.8B
• Weights: about 28.8 GB in F8_E4M3; about 14.4 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 4,489
• Likes: 242
• License: apache-2.0
• Quantized versions: 12
• Fine-tunes: 2
• Tags: transformers, safetensors, raon1_mini, text-generation, raon, conversational, mxfp8

Links

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

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