Raon-1 Nano: Krafton's Tiny Game-NPC LLM, Explained
Krafton distilled a 29B MoE into a 1.5B dense model called Raon-1 Nano for game NPCs. Here's how it works and how to run it locally.

What is Raon-1 Nano?
Raon-1 Nano is a 1.5 billion parameter open-weight language model released by Krafton, the South Korean studio best known for PUBG: Battlegrounds. It’s not trained from scratch in the usual sense. Krafton took its own 29 billion parameter mixture-of-experts (MoE) model, Raon-1 Mini, and converted it into a small dense model before continuing training on it. The result is released under an Apache 2.0 license on Hugging Face, with GGUF-style local deployment in mind, aimed squarely at running game NPCs and AI teammates on a player’s own machine rather than in the cloud.
TL;DR
- Krafton, the maker of PUBG, built Raon-1 Nano as part of a broader in-house model family called Raon, which already includes speech, TTS, and vision-encoder models.
- The model is a distilled dense model, converted from a 29B mixture-of-experts parent (Raon-1 Mini, 4.3B active parameters, 128 experts) down to a 1.5B dense architecture, then trained further with 3 trillion tokens, supervised fine-tuning, on-policy distillation, and reinforcement learning.
- Krafton published the conversion method as a NeurIPS 2025 oral paper, describing a diversity-aware expert scoring and grouping technique that beat standard pruning by 6.3 points on average while training 1.6x faster.
- The model is intentionally weak on general world knowledge (it trails on MMLU Pro) but strong on instruction following and on knowing what it doesn’t know, which matters more for an NPC than trivia recall.
- Licensing is Apache 2.0, and Krafton briefly gated the release before ungating it the next day after community feedback.
- Deployment targets are consumer-grade: an FP4 version fits on a single RTX 5090, and Krafton also shipped a version for Nvidia DGX Spark hardware.
- The practical use case lines up with Krafton’s shipped products: Smart Zoy (on-device NPCs in Krafton’s life-sim game inZOI) and PUBG Ally, an AI teammate that follows player instructions inside a match.
Build an onboarding app for 4,000 new hires.
Build ↵Why would a game studio build its own LLM?
Running a cloud-hosted LLM for every NPC conversation in a live game doesn’t scale economically. Latency matters, session volume is huge, and margins in games are thin compared to API costs. Krafton’s answer is to push inference onto the player’s own GPU, alongside the game itself. That means the model has to be small enough to share VRAM with the game, fast enough to respond without breaking immersion, and reliable enough to stay in character over a long session.
Krafton has already shipped on-device AI characters: Smart Zoy, NPCs in its life-simulation game inZOI, and PUBG Ally, an AI teammate that follows player callouts during a match (first demoed publicly by Nvidia in early 2025, later shipped in the live game). Krafton’s chief AI officer has described Raon-1 Nano as a better base model for these edge AI applications. Krafton hasn’t explicitly confirmed that PUBG Ally runs on Raon-1 Nano, but the design goals overlap closely enough that the connection is hard to miss.
An AI teammate in a shooter needs to track short-term context (what weapon you said you hate, what strategy you prefer), stay in tone, and use live game state (inventory, map zone, squad position) fed into its context window. It does not need to know world history or general trivia. That’s a very different profile from a general-purpose chatbot, and it explains why Krafton optimized for instruction-following and self-awareness of knowledge gaps rather than broad factual recall.
How was Raon-1 Nano distilled from a 29B MoE model?
The starting point is Raon-1 Mini, a 29 billion parameter mixture-of-experts model (128 experts, 6 active per token plus one shared expert, 4.3 billion active parameters per token) that Krafton trained from scratch rather than fine-tuning from an existing base like Qwen or Llama.
Most prior MoE compression work tries to shrink a big MoE into a smaller MoE. Krafton instead converted the MoE into a plain dense model. According to the paper, the process works in four steps:
- Score every expert to determine which ones matter most, using a custom “diversity-aware” scoring method that tries to preserve experts doing genuinely different things rather than redundant ones.
- Select and group the highest-value experts.
- Merge each group into a single dense feed-forward layer.
- Distill from the original MoE so the new dense model learns to match its parent’s outputs.
Krafton says it tested 350 different configurations of this pipeline and found expert scoring to be the variable that mattered most. Against a baseline of simply pruning a dense model to the same size, the diversity-aware approach beat it by 6.3 points on average while also training 1.6 times faster.
The raw numbers look like a 20x parameter reduction (29B to 1.5B), which is unusually aggressive for distillation. But since only 4.3B parameters activate per token in the MoE, the effective compute gap is smaller than the headline suggests. It’s closer to compressing a dynamic 4B model that reassembles itself per token into a fixed 1.5B dense model.
The conversion itself isn’t the final model. Krafton’s own paper notes that models straight out of this conversion step still trail Qwen 3 1.7B. To get to the released Nano model, Krafton added roughly 3 trillion tokens of further pretraining, supervised fine-tuning, on-policy distillation from the Mini model, and a reinforcement learning stage. Notably, no outside base model appears anywhere in this lineage: Mini and Nano are both trained in-house.
Why is a model that’s bad at trivia still useful?
On MMLU Pro, a broad knowledge benchmark, Raon-1 Mini finished last among the five models Krafton compared it to (76.5 versus 85.9 for Qwen 3 36B). That looks like a weakness until you look at what the model is for. On instruction-following benchmarks, Mini ranked first in Krafton’s comparison. On the Artificial Analysis Omniscience benchmark, which penalizes confidently wrong answers and rewards a model for correctly recognizing what it doesn’t know, it also came out on top.
That combination suggests Krafton isn’t trying to build a model that knows everything. It’s building one that knows the boundary of its own knowledge and defers to external context (game state, scripted lore, retrieved documents) rather than hallucinating facts. Feed forward layers in a transformer are generally thought to store a large share of a model’s factual knowledge; in an MoE, that’s the experts. Converting most experts away while preserving the attention layers plausibly strips out stored trivia faster than it strips out the ability to follow instructions and reason over supplied context. Krafton doesn’t make this claim explicitly in the paper, but it fits both the benchmark pattern and the NPC use case.
This also echoes an idea Andrej Karpathy has described as a “cognitive core”: a small model, on the order of a billion parameters, that handles reasoning and instruction-following well but relies on retrieval for facts rather than memorizing the open internet. Raon-1 Nano wasn’t necessarily built to chase that idea specifically, but the overlap with a game studio’s NPC requirements is obvious.
How do you run Raon-1 Nano locally?
Krafton published Raon-1 Nano weights on Hugging Face under the KRAFTON organization, with safetensors files, a custom model implementation, tokenizer files, and a separate MTP (multi-token prediction) component included in the repository. The license is Apache 2.0, meaning commercial use, modification, and redistribution are permitted without royalty.
For hardware, Krafton has released an FP4-quantized version small enough to run entirely on a single Nvidia RTX 5090, and a build targeting Nvidia’s DGX Spark desktop AI systems. Given the model’s size (1.5B parameters) and the availability of quantized formats, it’s also realistic to run on more modest consumer GPUs once community GGUF conversions circulate, similar to how other sub-2B models get packaged for tools like llama.cpp or Ollama. For the larger Raon-1 Mini (29B MoE, 4.3B active), Krafton currently offers a free hosted API for evaluation, alongside the option to self-host.
Because the model ships with Apache 2.0 licensing and open weights, it’s usable by developers outside Krafton’s own games: anyone building a narrow, context-heavy local agent (customer support kiosks, embedded assistants, offline tools) faces similar constraints to a game NPC: tight latency, limited memory, and a need to use supplied context well rather than recalling broad world knowledge.
Frequently Asked Questions
What is Raon-1 Nano used for?

It’s designed for on-device game NPCs and AI companions, the kind of role Krafton’s Smart Zoy characters and PUBG Ally teammate fill, where the model needs to follow instructions, track short-term context, and stay in character rather than answer general knowledge questions.
How is Raon-1 Nano different from a normal distilled model?
Most distillation goes from a large dense model to a smaller dense model, or from a large MoE to a smaller MoE. Krafton instead converted a 29B mixture-of-experts model (Raon-1 Mini) directly into a 1.5B dense model by scoring, grouping, and merging experts into standard feed-forward layers, then distilling further.
Is Raon-1 Nano good at general knowledge or coding?
No, and that appears to be by design. It underperforms larger models on broad benchmarks like MMLU Pro and isn’t built for agentic coding tasks. It scores well on instruction-following and on recognizing the limits of its own knowledge, which matters more for its intended NPC use case.
Can I run Raon-1 Nano on my own hardware?
Yes. Krafton released the weights under Apache 2.0 on Hugging Face along with an FP4 version that fits on a single RTX 5090, plus support for Nvidia DGX Spark hardware. Its small size also makes it a candidate for further community quantization for lower-end GPUs.
What license does Raon-1 Nano use?
Apache 2.0, a permissive open-source license that allows commercial use and modification. Krafton initially gated access to the model but removed the gate the following day after community feedback.


