Blog
RAON-1 MiniRAON-1 Nanorun locally

How to Run RAON-1 Mini and Nano Locally: Hardware and Setup

Krafton's RAON-1 Mini is a 29B MoE with 4.3B active, Apache 2.0. Here is the hardware, quantized builds and steps to run it and the 1.5B Nano locally.

Portrait of Luis Chavez-Mattos
Edited by Luis Chavez-Mattos, Director, Product··RSS

For Engineering and IT

Three comic panels show a young man with glasses amazed, then squeezing a large graphics card between his hands, then a close-up of the card with a small glowing light.

Can you run RAON-1 Mini and Nano on your own computer?

Yes. RAON-1 Mini is an open model from KRAFTON AI (the games company behind PUBG) with about 29 billion parameters in total, of which 4.3 billion are active for any one token. A 4-bit quantized build fits on a single 32 GB RTX 5090, and a separate GGUF release lets you run it through llama.cpp. RAON-1 Nano is a much smaller 1.5 billion parameter model distilled from Mini. Mini is released under the Apache 2.0 license, so you can use it commercially.

TL;DR

  • RAON-1 Mini is a 28.8B-parameter mixture-of-experts model with 4.3B active per token, trained from scratch by KRAFTON AI, not fine-tuned from someone else's model.
  • The 4-bit build is the one for consumer hardware. The model card says it enables inference on a single 32 GB RTX 5090, and the release also includes a DGX Spark version.
  • A GGUF release exists for llama.cpp, which is the most common way to run open models on a desktop or laptop.
  • It is a preview, not a final model. Specs and results may change.
  • The license is Apache 2.0, which allows commercial use, modification and redistribution with attribution terms.
  • Mini is weak on general knowledge and strong on following instructions. It is not built to be a coding model.
  • Nano (1.5B dense) is the small option for laptops and edge devices, aimed at tasks where you feed the model the facts it needs.

What are RAON-1 Mini and Nano?

KRAFTON AI is the AI group inside KRAFTON, the South Korean company that makes PUBG: Battlegrounds. It has been releasing open models under the Raon name, starting with speech and text-to-speech models and a vision encoder before reaching language models.

Mini is a mixture-of-experts (MoE) model. Instead of running every parameter for every word, an MoE routes each token through a small subset of specialist sub-networks called experts. Mini has 128 routed experts, uses 6 of them per token plus 1 shared expert, and so activates 4.3B of its 28.8B parameters. It supports up to 262,144 tokens of context, has built-in thinking, and supports tool calling for agent workflows (the model card names Hermes Agent, Pi and OpenCode).

Nano is a 1.5B dense model. Dense means every parameter is used for every token, the way most small models work. KRAFTON did not train it the usual way. According to the video coverage of the release, the team converted the Mini mixture-of-experts model into a dense one, then trained it further (3 trillion more tokens, supervised fine-tuning, distillation from Mini, and reinforcement learning). The conversion method has a paper accepted as an oral presentation at NeurIPS.

What hardware do you need to run RAON-1 Mini?

The key idea: an MoE model is fast for its size, but you still have to hold all of it in memory. The 4.3B active parameters decide how quickly it generates text. The 28.8B total decides how much memory you need.

These sizes are rough estimates from the parameter count (weights only, before the context cache and overhead):

BuildApprox. weight sizeRealistic hardware
BF16 (full precision)about 58 GBMulti-GPU workstation or large unified-memory machine
MXFP8 (main repo)about 29 GBTight on a 32 GB card once context is added; comfortable on larger memory
4-bit (NVFP4)about 15 GBA single RTX 5090 (32 GB), per the model card
GGUF Q6_K and smallervaries by quantDesktop GPUs, or a CPU plus RAM with patience

Two details help with planning:

  • Long context is cheap here. Only 7 of Mini's 25 layers keep a KV cache (the memory a model uses to remember the conversation), at about 8 KiB per token in BF16. At the full 262,144 tokens that works out to roughly 2 GiB. Many models of this size need far more.
  • Unified-memory machines are an option. According to the video coverage of the release, there is a version for DGX Sparks, and the model card's GGUF files target llama.cpp, which runs on a wide range of hardware. Check the repo for the specific files before you download tens of gigabytes.

Which quantized build should you pick?

Quantization shrinks a model by storing its numbers with fewer bits. You lose a little quality and gain a lot of memory.

A note on naming, because it trips people up. Coverage of the release described the single-5090 build as MXFP4. The model card lists the options as MXFP8 and NVFP4, and says NVFP4 is the one that enables inference on a single RTX 5090. Both are 4-bit-class and 8-bit-class formats from the same family of low-precision formats. Go by the file names in the Hugging Face repo, not by any blog's label, mine included.

A simple way to choose:

  1. You have a 32 GB RTX 5090: use the NVFP4 build.
  2. You have a DGX Spark: use the build released for it.
  3. You have a smaller GPU, a laptop or a Mac: use the GGUF repo (KRAFTON/Raon-1-mini-preview-gguf). It was converted from the BF16 weights, and Q6_K and smaller quants were made with an importance matrix, a technique that protects the most sensitive weights when compressing. Pick the largest quant that fits in your memory with a few GB spare.
  4. You have no suitable hardware: use the free playground linked from the model card. It is available for a limited time.

How do you run it with llama.cpp?

llama.cpp is a free program that runs open models on ordinary computers. The steps below are the general path. Exact file names and any required version are on the model card.

  1. Get a current llama.cpp build. Mini uses a newer hybrid attention design, so an old build may not recognize it. Update before troubleshooting anything else.
  2. Download one GGUF quant from the KRAFTON GGUF repo on Hugging Face. Choose by the memory table above.
  3. Start a local server. A typical command looks like this (replace the path with your file):
llama-server -m /path/to/raon-1-mini-quant.gguf -c 32768
  1. Keep the context modest at first. The model supports 262,144 tokens, but you do not need that to test it. Start at 8K to 32K and raise it once things work.
  2. Open the local web page the server provides, or point any tool that speaks the OpenAI-style API at it.

If you want faster output, the model card says a DFlash2 draft model is provided for speculative decoding. That technique has a small model guess ahead while the large one verifies, which speeds up generation. Treat it as a second step after the basic setup works.

Is RAON-1 really Apache 2.0?

Yes. The model card lists Apache 2.0. Per the video coverage, KRAFTON first released it under a different, gated license and switched to Apache 2.0 the next day after feedback.

In plain terms, Apache 2.0 lets you use the model at work, build it into a product, modify it and ship it, as long as you keep the license and notices. It also includes a patent grant. If your company has a legal review process for open-source software, this is a license those teams already know, which makes approval easier than with custom "open" licenses.

How good is RAON-1 Mini, and what is it bad at?

The benchmark table on the model card compares Mini with Qwen3.6-35B-A3B, Gemma 4 26B-A4B, Qwen3.8-27B and Muse Glimmer 30B. The pattern is clear:

BenchmarkRaon-1 mini previewQwen3.6-35B-A3BGemma 4 26B-A4BQwen3.8-27BMuse Glimmer 30B
MMLU Pro (general knowledge)76.585.982.985.784.0
AA-Omniscience3.5-7.2-46.6-0.5-34.7

Mini scores lowest on MMLU Pro, a broad knowledge test. But AA-Omniscience penalizes confident wrong answers and rewards saying "I don't know," and Mini scores highest there. The model card also says it leads the comparison models on its instruction-following and search benchmarks, and reports 84.3 on GPQA Diamond.

The reading: this is a model that knows less trivia but is more careful about what it doesn't know. That fits a model meant to work with information you hand it (documents, game state, search results). The video's host noted that Mini is also weak on terminal-based coding benchmarks, so do not expect it to be a coding agent. That interpretation of the numbers is the commentator's, not something KRAFTON has stated.

What is RAON-1 Nano good for?

At 1.5B parameters, weights take roughly 3 GB at 16-bit precision (an estimate from the parameter count), and less when quantized. That is small enough for a laptop, a modest GPU or an on-device deployment sharing hardware with other software.

KRAFTON's own use case explains the design. It ships AI characters in its games, and a game-side model has to be small, fast and share a GPU with the game. Such a model needs to follow instructions, keep a persona and use what is in its context, not recall world trivia. The same constraints apply to a kiosk assistant, an internal helper on a laptop, or a narrow support tool. For these, you supply the facts through the prompt or a document search, and the model does the reasoning and wording.

KRAFTON has not said which production features run on Nano, so treat "it powers PUBG Ally" as unconfirmed. For Nano's exact files and requirements, check KRAFTON's Hugging Face page.

What this means for you

If you are a curious non-developer: try the free playground first. It tells you in ten minutes whether the model's style suits you, with nothing installed.

If you work in IT or data teams: the combination of Apache 2.0, a single-GPU build and a GGUF release makes Mini easy to pilot on hardware you may already own. Test it on your own documents with retrieval, since its strength is working with provided information, not recalling facts. Remember it is labeled a preview.

If you handle sensitive data (legal, HR, finance): a local model keeps text on your machine. That is the main reason to bother. Confirm with your security team before pointing it at real records.

If you want a general chatbot or coding assistant: this is probably not the right pick. The knowledge score is the lowest in its comparison group and it is not built for coding.

Can you build a tool around a local model like this?

A local model is only useful once something sits in front of it, such as a request form, a document search page or a tracker with a login. With Remy, you describe that app in plain English and it builds and deploys it, including database and sign-in. Remy's model settings let you register custom or self-hosted models, or reach local models over a secure tunnel. A reasonable first project is an internal question-and-answer page over your team's documents that you can point at a local model while you evaluate it, instead of paying for another subscription.

Frequently Asked Questions

4 questions

Can RAON-1 Mini run on a laptop?

Possibly, through the GGUF quants and llama.cpp, if the laptop has enough memory for the quant you pick. The model card's confirmed single-GPU path is a 32 GB RTX 5090. Smaller quants trade quality for size, so expect slower, lower-quality results on laptops.

Is RAON-1 Mini free for commercial use?

The model card lists the Apache 2.0 license, which permits commercial use. Keep the license text and notices with anything you redistribute, and check the repository for the current terms.

Why does a 29B model need less compute than its size suggests?

Because it is a mixture-of-experts model. Only 4.3B parameters are active for each token, so generation is faster than for a dense 29B model. Memory needs are still based on the full 28.8B, since all experts must be loaded.

Is RAON-1 Mini a finished model?

No. The model card calls it a preview release and says specs and results may change before the final version.

How we source, write, and correct our articles. Editorial standards

Related articles

Straight to your inbox.
New pieces as they publish. No noise.