Sam Witteveen · 2026-11-17 · notable
Sam Witteveen — 'RAON-1 Nano: Can a Tiny LLM Power Game NPCs?'
In short
Sam Witteveen tests KRAFTON's Raon-1 nano, a small Apache-2.0 model distilled from the larger Raon-1 mini, and asks whether a model this small can run the characters in a video game.
Sam Witteveen · 2026-11-17 · 2 sources checked

Sam Witteveen tries KRAFTON's tiny Raon-1 nano model as the brain for game characters.
What is it?
'RAON-1 Nano: Can a Tiny LLM Power Game NPCs?' went up on Sam Witteveen's channel on 17 November 2026. The video looks at Raon-1 nano from KRAFTON, the games company, and why a games studio builds its own language models. It ends with a live demo and takeaways.
How does it work?
According to KRAFTON's model card, Raon-1 nano is a dense Korean and English model with 1.5B effective parameters, made from the 30B-A4B Raon-1 mini preview mixture-of-experts model by pruning it into a dense model and then distilling it. The video walks through that MoE-to-dense step, the benchmarks, the parent model's free API and how to run the models locally.
Why does it matter?
Game characters need a model that answers quickly on a player's own machine, not one that calls a large cloud model for every line. A 1.5B-parameter model with an Apache 2.0 license and builds for GGUF, MLX and LiteRT-LM is the kind of model a studio can ship inside a game.
Who is it for?
game developers and local-AI builders
Try it
huggingface.co/KRAFTON/Raon-1-nano