Raon-1 nano

Raon-1 nano

Generates text and follows instructions in Korean and English, keeps conversation context and calls tools. A 3.1B-parameter language model with a 262,144-token context, for developers and on-device apps.

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: KRAFTON AI built Raon-1 nano as a compact language model. The card counts 1.5B parameters as effective and 3.1B overall, because a per-layer embedding table adds 1.5B. It follows instructions, keeps track of a conversation and picks the right tool to call, in Korean and English. It belongs to the Raon open model family.

Who it suits: It suits developers who want a small Korean and English model that can run on a device, and researchers who study its architecture.

Title: Who might use it
Industry | Who uses it | How they use it | Example
Mobile apps | App developers | Run a small model on a device for chat and tool calls | Controlling phone functions through a conversation
Office software | Developers | Add instruction-following chat in Korean and English | An assistant that calls tools for everyday office tasks
Korean-language products | Product teams | Build Korean conversation features | A Korean chat assistant inside an app
AI research | Researchers | Study the hybrid attention design and the pruning method | Reading the architecture and training notes on the card

People who need image or audio input can skip it, since it is a text generation model.

Where to use it:

• Python with the transformers library: the model is published for it
• Local inference formats: GGUF, MLX and LiteRT-LM, including on Apple silicon
• Windows, macOS or Linux: download the files with Python

Title: Jobs it does
Use case | Who | What they get back
Conversation | Developers | Replies in Korean or English that stay in context
Instruction following | Product teams | Text written to match the instruction given
Tool calling | App builders | A choice of tool inside a conversation

How to set it up: You need these first:

• Python on Windows, macOS or Linux
• Storage for about 7.1 GB of weights in BF16, or an estimated 1.8 GB at 4-bit

1. Install the Hugging Face hub package.
2. Download the model files.
3. Follow the run steps on the model card, which cover loading and prompting.

pip install -U huggingface_hub
hf download KRAFTON/Raon-1-nano

How to use it: After the download, load the model with transformers or use one of the on-device formats, then send it a prompt in Korean or English. This command fetches the files.

hf download KRAFTON/Raon-1-nano

The model answers with text. In a conversation it can call tools, which the card ties to everyday jobs such as office work and mobile control. The card holds the exact loading steps.

Example: You ask it in Korean to follow a short instruction, such as rewriting a note more politely, and keep chatting about it. It replies in text and keeps the earlier turns in mind.

For newcomers and experts: If you are new to this, treat it as a small chat model that speaks Korean and English. Download it, then try one short instruction and see how it replies.

If you work in the field, the card describes a hybrid stack of 24 layers, where 18 use linear attention and the other 6 use full attention. It adds a small embedding in every layer and ties the output head to the input embedding. Just those 6 layers store a KV cache, which costs roughly 7 KiB for each token in BF16. It was pruned from Raon-1 mini preview, then trained on 3T more tokens, followed by supervised fine-tuning, on-policy distillation and reinforcement learning. The card says it beats the models it compares against in tests of following instructions, keeping context and talking in Korean.

Download the weights
Load them with transformers or an on-device format
Send a Korean or English prompt
Get text back or a tool call

Good to know: The card lists a context length of 262,144 tokens and a vocabulary of 259,840 entries.

Title: Weight size in GB
BF16: 7.1 GB
4-bit estimate: 1.8 GB

• License: apache-2.0
• Cost to run: none named
• Data it reaches: only the prompts you give it, with no outside service named

On Hugging Face, 13 November 2026

• Task: text generation
• Library: transformers
• Parameters: 3.5B
• Weights: about 7.1 GB in BF16; about 1.8 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 446
• Likes: 77
• License: apache-2.0
• Quantized versions: 5
• Fine-tunes: 1
• Tags: transformers, safetensors, raon1_nano, text-generation, raon, conversational

Links

• Model: https://huggingface.co/KRAFTON/Raon-1-nano
• Files: https://huggingface.co/KRAFTON/Raon-1-nano/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download KRAFTON/Raon-1-nano

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