Raon-1 mini: KRAFTON's Apache-licensed 30B MoE model for Korean and agents

Game company KRAFTON has released Raon-1 mini (Preview), a language model it trained itself, with 28.8B parameters, of which only 4.3B are active per token. It's said to be especially strong in Korean, reasoning and agent tasks, and runs in NVFP4 on an RTX 5090.

Model releases Try it yourself Oct 21, 2026 4 min read Source: Hugging Face Trending

What this is about

KRAFTON AI, the AI division of game company KRAFTON, has released Raon-1 mini preview on Hugging Face. It's a Mixture-of-Experts (MoE) language model, meaning it consists of many specialized sub-networks ("experts"), only a few of which are active for each word. Raon-1 mini has 28.8 billion parameters in total (the adjustable values learned during training), but only 4.3 billion are used per token, hence the name "30B-A4B". According to KRAFTON, the model was trained entirely in-house from scratch and is released under the permissive Apache 2.0 license. It is explicitly a preview: specifications and results may change before the final version. There's also a free, time-limited playground you can try.

Architecture

Raon-1 mini uses a hybrid 25-layer design: 18 are linear attention layers (Raon Delta Attention, RDA) and 7 are full attention layers (Gated MLA).

Raon Delta Attention (RDA): Decouples the erase and write gates of Kimi Delta Attention, which the maker says improves long-context performance compared with KDA. These layers have a fixed-size memory.
Small KV cache: Only the 7 MLA layers keep a KV cache (a buffer for text already read), at about 8 KiB per token in BF16. This makes long contexts easier on memory.
Raon-mHC: A variant of DeepSeek's Manifold-Constrained Hyper-Connections with lower compute cost, which speeds up training.
Sparse MoE: 6 of 128 experts plus 1 shared expert are active per token.
Context: Natively up to 262,144 tokens. There is no positional encoding (NoPE).

Other specs: hidden size 2,560, 40 attention heads, vocabulary of 259,840, SiTU-GLU as the activation function, untied LM head.

Capabilities

Reasoning: Built-in "thinking" before answering.
Multilingual: Trained on Korean, English, Japanese, Chinese, Spanish, French and Thai, with particular strength in Korean.
Agents: Tool calling for agent workflows, with support for Hermes Agent, Pi and OpenCode.
Speculative decoding: A DFlash2 draft model speeds up text output.
Quantization: MXFP8 and NVFP4 are available. NVFP4 allows inference on a single RTX 5090 with 32 GB VRAM.

Benchmarks (maker's figures)

The comparison covers MoE models (Qwen3.6-35B-A3B, Gemma 4 26B-A4B) and dense models (Qwen3.8-27B, Muse Glimmer 30B). The table shows an excerpt:

Benchmark	Raon-1 mini	Qwen3.6-35B-A3B	Gemma 4 26B-A4B	Qwen3.8-27B	Muse Glimmer 30B
MMLU Pro	76.5	85.9	82.9	85.7	84.0
GPQA Diamond	84.3	83.2	80.7	89.7	82.6
AIME 2026	93.3	92.9	87.1	98.8	92.9
HMMT Feb 2026	84.1	84.5	79.9	95.8	74.6
IFBench	84.2	66.6	80.9	82.4	78.7
MultiChallenge	59.0	58.2	58.4	63.6	63.4
LiveCodeBench v6	81.6	81.2	83.4	79.3	77.4
SWE-Verified	70.8	63.2	61.6	74.6	72.0
Terminal-Bench Lite	47.0	63.3	52.0	67.0	70.7
BrowseComp Plus	55.7	34.8	32.2	58.0	52.8

The picture is mixed. Raon-1 mini leads the MoE models in reasoning, instruction following and search, but trails clearly on knowledge (MMLU Pro) and Terminal-Bench Lite. The table in the source is cut off, so further benchmarks aren't covered here.

What you need to try it

According to the maker, the NVFP4 version runs on a single Nvidia RTX 5090 (32 GB VRAM). The larger BF16 or MXFP8 weights need considerably more video memory (28.8B parameters in BF16 alone come to roughly 58 GB). The card mentions Transformers with custom_code; details on Mac, CPU or Colab aren't in the source. No GGUF version is listed. Plan for a few dozen gigabytes of disk space. If you'd rather not install anything, you can use the free, time-limited playground.

Our take

Raon-1 mini is an interesting open-weight model from South Korea, and it shows that game companies are training their own foundation models. The combination of few active parameters, a small KV cache and 262k context is attractive for local use. Keep in mind that all benchmarks come from the maker, the preview version may still change, and the model trails competitors on knowledge and terminal tasks. The stronger dense models in the comparison are in part better. Strong Korean performance is claimed, but the tables we captured contain no direct evidence for it. With only 249 downloads at the time of capture, community experience is still thin.

Terms explained simply

Mixture-of-Experts (MoE): A model made of many specialized sub-networks, only some of which are active per token. This saves compute while keeping a large total capacity.
Active parameters: The share of total parameters that actually does the computing for each token. Here, 4.3B of 28.8B.
KV cache: A buffer for text that has already been processed. It grows with context length and eats up VRAM.
Linear attention: A variant of attention with fixed memory that handles long texts more cheaply than classic full attention.
Speculative decoding: A small draft model guesses words ahead and the large model checks them, which speeds up output.
NVFP4 / MXFP8: Space-saving 4-bit and 8-bit number formats that sharply cut the model's memory needs.
Context length: How much text the model can consider at once. Here, up to 262,144 tokens.

Sources

Hugging Face Trending huggingface.co
Google News (KR) news.google.com
Google News (KR) news.google.com

#krafton#raon#moe#korean#apache-2.0#agents

Read the original source

Try it yourself: Raon-1-mini-preview

Explained step by step, no prior knowledge needed – on your own computer or free in your browser.

What can I do with it?

Raon-1 mini is a chat and reasoning model that is especially good at Korean, can use tools (tool calling) and handles very long texts of up to 262k tokens. You can run it locally as an assistant or an agent backend.

Who is it worth it for?

For enthusiasts with a powerful Nvidia card (e.g. an RTX 5090) and anyone who wants to test Korean text or agent workflows. If you only have a small GPU, you're better off waiting for GGUF versions.

Size28.8 B parameters (4.3 billion active)

Download20 GB (about 26 min at 100 Mbit/s)

Licenseapache-2.0

TypeLanguage model (text chat)

Hugging FaceKRAFTON/Raon-1-mini-preview

What your computer needs

Graphics card (NVIDIA), fully in VRAM at least 23 GB VRAMe.g. RTX 3090, RTX 4090, Tesla P40 – fastest; with less VRAM, see the next line

Graphics card + system memory (MoE offloading) from approx. 6 GB VRAM + 32 GB RAMdie Experten liegen im Arbeitsspeicher, nur der Kern auf der Grafikkarte (llama.cpp/LM Studio). Langsamer als komplett im VRAM, aber auf normalen PCs machbar

Mac (Apple Silicon M1–M5) from 36 GB unified memoryThe GPU and CPU share the memory – the model can use about 70% of it

CPU only (no graphics card) Runs smoothly without a graphics cardan estimated 9 words per second on a current desktop PC (MoE: only 4.3 billion parameters active per word). This requires at least 32 GB of RAM.

Storage approx. 22 GB freeModel download plus programs

Google Colab Colab Pro onlywith MoE offloading on a T4/L4 plus a High-RAM runtime (about 51 GB of RAM); free Colab doesn't have enough RAM

Which variant is right for me?

Many models come in several sizes or "quantizations." These are compressed versions: smaller and faster, but slightly less accurate. Rule of thumb: pick the largest variant whose memory requirement fits in your VRAM (video memory). The scripts below choose this for you automatically.

Variant	Download	Needs approx.	Compatible graphics cards
`Q4_K_M` Default	20 GB	23 GB	e.g. RTX 3090, RTX 4090, Tesla P40
`Q5_K_M`	22 GB	26 GB	e.g. RTX 5090
`Q6_K`	26 GB	30 GB	e.g. RTX 5090
`Q8_0`	31 GB	36 GB	e.g. 2× RTX 3090/4090 or RTX A6000

The NVFP4 variant is meant for single GPUs (it runs on an RTX 5090 with 32 GB, according to the maker). MXFP8 and BF16 need considerably more memory.

According to the maker, NVFP4 runs on a single RTX 5090 with 32 GB VRAM.
Only 4.3B parameters are active per token, but all 28.8B must fit in memory.
The KV cache is small (about 8 KiB per token in BF16), so long contexts are cheaper.
A DFlash2 draft model enables faster speculative decoding.

Gated access: You need a free Hugging Face account. Log in on the model page and accept the license. Then create an access token (type Read) at huggingface.co/settings/tokens and run hf auth login once before you start.

How do I find out what my computer has?

Windows: Press Ctrl + Shift + Esc → Task Manager → Performance tab → click GPU on the left. At the bottom you'll see Dedicated GPU memory (that's your VRAM). You'll find system memory under Memory.
Mac: Apple menu at the top left → About This Mac. You'll see your chip (e.g. M2) and memory (e.g. 16 GB). On Apple chips, graphics and processor share this memory.
Linux: Type nvidia-smi in the terminal (shows your graphics card and its memory) and free -h for system memory.

Guide

Google Colab (in your browser, no installation)

Google Colab is a free service from Google. It gives you a computer with a graphics card in your browser for a few hours. All you need is a Google account. For this model: Colab Pro only – needs an A100 with 40 GB.

Download the notebook. Click the button and save the file.
Download colab.ipynb
Open Colab. Go to colab.research.google.com and sign in with your Google account.
Upload the notebook. In the window that appears, click Upload on the left (or choose File → Upload notebook at the top) and select the file colab.ipynb.
Turn on the GPU. In the top menu, go to Runtime → Change runtime type → under Hardware accelerator, select T4 GPU (with Colab Pro: L4 or A100) → Save.
Start everything. Runtime → Run all (keyboard shortcut Ctrl + F9). If you see the warning “This notebook was not authored by Google,” click Run anyway.
Wait. The cells run one after another. A green checkmark appears to the left of each cell when it's done. Installation and download take about 11 minutes here.
Try it out. The result appears below the last cell. To test something else, change the text in the cell and run it again with Shift + Enter. The last cell starts a proper chat with an input field.
Clean up. When you're done: Runtime → Disconnect and delete runtime. This saves your free quota.

Good to know: The free GPU is limited (usually a few hours per day). After about 90 minutes of inactivity, Colab disconnects, and you'll have to start everything over.

Windows 10 / 11

The easiest way is Ollama, a free program that launches AI models with a single command.

Update your graphics driver (NVIDIA cards only): Install the NVIDIA app or the latest driver. For AMD cards, use the latest Adrenalin driver.
Install Ollama. Download the file OllamaSetup.exe from ollama.com/download/windows, double-click it, and choose Install. Afterward, a small llama icon appears in the taskbar at the bottom right.
Open a terminal window. Press the Windows key, type PowerShell and press Enter. A blue or black window opens.
Start the model. Copy this command ("Copy" button), paste it into the window with a right-click, and press Enter:
```
ollama run hf.co/KRAFTON/Raon-1-mini-preview-gguf:Q4_K_M
```
The first time, Ollama downloads the model (20 GB (about 26 min at 100 Mbit/s)). After that, it's instant.
Start typing. Once >>> Send a message appears, type your question and press Enter. Type /bye to end the chat.

Not enough memory? Pick a smaller variant from the table above, e.g. ollama run hf.co/KRAFTON/Raon-1-mini-preview-gguf:Q4_K_M.

Prefer automatic? A script that picks the right variant for you

Download both files into the same folder: try_windows.ps1 demo.py
Open the folder in File Explorer, right-click an empty spot and choose Open in Terminal.

Paste this command and press Enter:

powershell -ExecutionPolicy Bypass -File try_windows.ps1

If needed, the script installs Ollama via winget, measures your VRAM (video memory) and downloads the matching variant.

macOS

Macs with an Apple chip (M1 to M5) are ideal. Older Intel Macs are very slow.

Install Ollama. Download it from ollama.com/download/mac, open the file, and drag Ollama into your Applications folder. Launch Ollama once – a llama appears in the menu bar at the top.
Open Terminal. Press Cmd + Space, type Terminal, and press Enter.
Start the model. Paste the command (Cmd + V) and press Enter:
```
ollama run hf.co/KRAFTON/Raon-1-mini-preview-gguf:Q4_K_M
```
The first time, the model is downloaded (20 GB (about 26 min at 100 Mbit/s)).
Start typing. Type your question and press Enter. Quit with /bye.

Pick the right variant automatically

try_mac.sh

cd ~/Downloads
bash try_mac.sh

Linux

Check your GPU. nvidia-smi should show your card. If it doesn't, install the driver on Ubuntu and restart:
```
sudo ubuntu-drivers install
```

Install Ollama:

curl -fsSL https://ollama.com/install.sh | sh

Start the model:
```
ollama run hf.co/KRAFTON/Raon-1-mini-preview-gguf:Q4_K_M
```
Exit with /bye. Use ollama ps to check whether the model is running on the GPU.

Or let it pick the right variant automatically:

try_linux.sh

bash try_linux.sh
# mit Python/transformers statt Ollama:
MODE=hf bash try_linux.sh

Try it

Explain step by step how to calculate the sum of the first 100 square numbers.

Translate this paragraph into Korean and briefly explain the politeness levels: 'Thanks for your help yesterday.'

Write a Python script that merges all CSV files in a folder, and explain the code.

If something goes wrong

The model uses custom_code, so you must allow trust_remote_code when loading it with Transformers.

It's a preview, so results may change in the final version.

According to the source, no GGUF/Ollama version is available. NVFP4 requires compatible Nvidia hardware.

For a quick test without installing anything, you can use the free playground.

“out of memory,” a crash, or your computer freezes

The model is too big for your VRAM (video memory). Pick a smaller variant from the table above and close other programs, especially games and browsers with lots of tabs.

It runs, but extremely slowly

Your CPU is probably doing the work instead of your graphics card. In Ollama, the PROCESSOR column of ollama ps shows how much runs on the GPU. If it says "100% CPU", update your graphics driver or pick a smaller variant that fits entirely into VRAM (video memory).

“ollama” or “python” is not found

Close the terminal window and open it again so it recognizes the newly installed program. For Python on Windows: reinstall it and check Add python.exe to PATH.

Windows: "Running scripts is disabled on this system"

Run the script with exactly the command above: powershell -ExecutionPolicy Bypass -File try_windows.ps1. This allows execution for this one time only.

Error message when loading the model (version, "trust_remote_code", unknown architecture)

Brand-new models often need the latest software version. With Ollama: update Ollama. With Python: pip install -U transformers in the activated environment. The Model Card usually lists the exact version.

Tokenstrom generates the scripts and notebook automatically. Please take a quick look before running them – they're plain, readable text. Mind the model's license: apache-2.0.

Ask the AI assistant

Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.

Connecting …

The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.

Comments

What do you think? Already tried it? Write it here, no sign-up needed.

Loading comments …

Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.