# Q&A: How KRAFTON Built PUBG Ally, a Co-Playable Character Powered by NVIDIA ACE

[NVIDIA Developer Blog](https://yomu.fyi/company/nvidia-developer-blog) · Elizabeth Goodman · Jun 25, 2026

**Type:** Explainer

## Summary

KRAFTON developed PUBG Ally, an on-device AI co-playable character for PUBG: BATTLEGROUNDS, using NVIDIA ACE tooling and models. Conventional non-playable characters rely on fixed dialogue, but PUBG Ally processes live game states and player voice input in real time. The interactive pipeline combines automatic speech recognition, a quantized Mistral-NeMo-Minitron-2B small language model, and a custom text-to-speech model running locally on client GPUs with at least 8GB of VRAM. To minimize latency, the architecture separates reactive gameplay into a behavior-tree System 1 layer while routing strategic decisions and speech generation to the language model. KRAFTON tuned the non-deterministic system across multi-layer evaluations and large-scale playtests with over a thousand players.

## Context

Traditional game AI companions and NPCs rely on fixed dialogue and struggle with real-time cooperation, context adaptation, and memory in dynamic, fast-paced battle royale environments.

## Approach / What changed

KRAFTON deployed an on-device pipeline via NVIDIA ACE featuring ASR, a quantized 2B-parameter SLM (Mistral-NeMo-Minitron-2B), and in-house TTS. They split operations into a fast behavior-tree layer (System 1) for tick-rate actions and an SLM reasoning layer (System 2) for dialogue and coordination, while optimizing KV caching and constraining domain scope.

## Takeaways

- PUBG Ally runs entirely on-device by utilizing a quantized Mistral-NeMo-Minitron-2B model that fits into graphics memory alongside the game on GPUs with 8GB of VRAM.
- The dual-layer architecture delegates tick-rate movement and combat reflexes to a behavior tree so gameplay reactions never stall on language model inference.
- Optimizing prompt structure to keep instructions stable across turns maximizes KV cache reuse and keeps on-device SLM inference latency predictable.

**Tags:** [Architecture](https://yomu.fyi/topic/architecture), [LLMs](https://yomu.fyi/topic/llm), [Machine Learning](https://yomu.fyi/topic/machine-learning), [Performance](https://yomu.fyi/topic/performance)

[Read original post](https://developer.nvidia.com/blog/how-krafton-built-pubg-ally-a-co-playable-character-powered-by-nvidia-ace)