KRAFTON has shipped PUBG Ally, an AI teammate for PUBG: BATTLEGROUNDS that listens to natural voice commands, reasons about the match, and talks and acts in real time. NVIDIA calls it a "co-playable character" (CPC), a new category that sits between a scripted NPC and a human squadmate. It is one of the first shipping games to run a language model fully on the player's own GPU.
What Happened
KRAFTON built Ally on NVIDIA ACE, the company's suite of on-device models for speech and reasoning in games. The character was first previewed alongside autonomous AI companions in PUBG and NARAKA: BLADEPOINT, and KRAFTON has now detailed the production architecture. Ally understands voice in English, Korean, and Chinese, retains memory across play sessions, and makes its own decisions about looting, repositioning, and reviving teammates rather than following a fixed script.
Why It Matters
Most AI characters in games today phone a cloud model on every line of dialogue, which adds latency and a per-message bill that does not scale to millions of players. Ally runs locally instead, so there is no network round trip and no inference cost per match. For game creators, this is the clearest proof yet that on-device character AI is production-ready, not a demo. It builds directly on the tooling we covered in our deep dive on the NVIDIA ACE Game Agent SDK.
Key Details
The reasoning brain is a quantized Mistral-NeMo-Minitron small language model (the 2B variant) that fits on GPUs with as little as 8GB of VRAM. KRAFTON paired it with an automatic speech recognition front end and a custom in-house text-to-speech model, all running on the device. The team split the workload with a System 1 / System 2 design: fast behavior trees handle reflex actions while the language model handles deliberate reasoning, so reaction time never waits on inference. Ally also stays grounded by reading live game state through tool calls, which keeps it factually accurate about weapons, items, and match conditions. Full model documentation lives on the NVIDIA ACE for Games page.
What to Do Next
If you build games and want to test an on-device companion of your own, start with NVIDIA's in-game inference SDK, which packages the same speech, reasoning, and animation models KRAFTON used. Try a narrow scope first: a single companion with a fixed toolset and a small quantized model, grounded in your engine's real game state. PUBG Ally shows the pattern works at scale when reflex actions and language reasoning are kept on separate tracks.
