RAON-1 Nano: The Tiny LLM for AI Game Characters

RAON-1 Nano: The Tiny LLM for AI Game Characters

More

Summary

Krafton, the South Korean company behind PUBG: Battlegrounds, has released RAON-1 Nano, an open 1.5 billion parameter language model built for AI game characters. This video from Sam Witteveen looks at what makes it different and why a games company wants to own its models.

Rather than training a small model from scratch, Krafton converted its own 30 billion parameter mixture-of-experts model into a compact dense model. The video explains the method from the team’s NeurIPS oral paper: scoring each expert, selecting and grouping them, merging each group into a dense feed-forward layer, and distilling from the parent model. Out of 350 configurations tested, the expert scoring approach mattered most, and the diversity-aware scoring beat pruning a dense model to the same size by 6.3 points on average on the authors’ benchmarks.

The video also covers the product context: on-device NPCs in the Sims-style game inZOI and the PUBG Ally AI teammate. A small model that follows instructions, remembers context, stays in character, and uses injected game state matters more here than broad world knowledge, and running locally avoids costly cloud inference. It also notes where the model is weak, such as agentic coding benchmarks like Terminal-Bench.


📺 Source: Sam Witteveen · Published November 17, 2026
🏷️ Format: Deep Dive

1 Item

Channels