KRAFTON adds Ally, a voice-driven AI teammate built on NVIDIA ACE, to PUBG

KRAFTON has added PUBG Ally, an AI teammate built on NVIDIA ACE, to PUBG: BATTLEGROUNDS. The character uses speech recognition to understand free-form voice commands, generates a tactical response with a 2-billion-parameter language model, and replies with a synthesized voice, all in real time. There are no prewritten lines: Ally responds to the specific situation on the battlefield.

AI-processed from NVIDIA Developer Blog; edited by Hamidun News
KRAFTON adds Ally, a voice-driven AI teammate built on NVIDIA ACE, to PUBG
Source: NVIDIA Developer Blog. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceNVIDIA Developer Blog ↗
Editor-in-Chief
Key metric #1Understands free-form voice commands with no predefined syntax
Key metric #2Makes tactical decisions based on the real-time game state
Key metric #3Adapts responses to the combat context instead of using canned lines

KRAFTON has added PUBG Ally to PUBG: BATTLEGROUNDS. It is a next-generation AI teammate that hears the player's voice commands, makes tactical decisions, and replies with a synthesized voice. It doesn't rely on prewritten lines or dialogue trees. Instead, the system responds to the specific situation in real time. Under the hood is the NVIDIA ACE stack.

How the model pipeline works

PUBG Ally runs on three specialized components that work in sequence.

The first is automatic speech recognition (ASR). The player speaks into the microphone: "Cover the right, I'm heading for the circle." ASR converts the audio to text and hands it to the language model.

The second component is a 2-billion-parameter SLM. It receives the player's request along with the current game state: the map, enemy positions, inventory, and the safe zone. From that data, the model produces a tactically sound response and decides what Ally should do.

The third component is text-to-speech: Ally speaks in a natural voice through the headset instead of posting text in chat. The entire pipeline completes within a few hundred milliseconds, so the delay goes unnoticed in the pace of a match.

Why a compact model

Choosing a 2-billion-parameter model was a deliberate engineering decision, not a forced compromise. Larger LLMs produce richer responses, but a shooter doesn't give you three seconds to wait. NVIDIA ACE was designed from the start for edge scenarios like this, with tight latency constraints.

The compact model fits in GPU video memory alongside the game's main rendering. Some of the computation runs directly on the client without calling a server. That removes the dependence on ping and reduces infrastructure load across millions of concurrent game sessions.

"We wanted

Ally to feel like a real teammate, not a chatbot inside the game," the KRAFTON team wrote on the NVIDIA Developer blog.

How Ally differs from classic NPCs

Traditional game companions run on scripted dialogue trees: developers write every line and every trigger condition in advance. That approach is predictable and controllable, but it breaks down in unusual scenarios, and players quickly sense the artificiality.

PUBG Ally isn't bound by a script. The language model generates a response for the specific moment: being surrounded in a late zone with no ammo calls for one tactic, an open field early in the match calls for another.

Key capabilities of the system:

  • Understands free-form voice commands with no predefined syntax
  • Makes tactical decisions based on the real-time game state
  • Adapts responses to the combat context instead of using canned lines
  • Delivers voice responses without latency that hurts gameplay
  • Runs locally on the player's GPU, with no server ping

What it means

PUBG Ally is one of the first voice-driven AI teammates in a mass-market online game. NVIDIA ACE gives other developers a ready-made stack for building similar systems. If KRAFTON's experiment proves commercially successful, voice-driven AI teammates with genuine context awareness will become a new standard in shooters, RPGs, and strategy games, and will permanently reshape what players expect from NPCs.

Story

Key Questions Answered

How the model pipeline works?

PUBG Ally runs on three specialized components that work in sequence. The first is automatic speech recognition (ASR). The player speaks into the microphone: "Cover the right, I'm heading for the circle." ASR converts the audio to text and hands it to the language model. The second component is a 2-billion-parameter SLM. It receives the player's request along with the current game state: the map, enemy positions, inventory, and the safe zone. From that data, the model produces a tactically sound response and decides what Ally should do. The third component is text-to-speech

Why a compact model?

Choosing a 2-billion-parameter model was a deliberate engineering decision, not a forced compromise. Larger LLMs produce richer responses, but a shooter doesn't give you three seconds to wait. NVIDIA ACE was designed from the start for edge scenarios like this, with tight latency constraints. The compact model fits in GPU video memory alongside the game's main rendering. Some of the computation runs directly on the client without calling a server. That removes the dependence on ping and reduces infrastructure load across millions of concurrent game sessions. "We wanted Ally

How Ally differs from classic NPCs?

Traditional game companions run on scripted dialogue trees: developers write every line and every trigger condition in advance. That approach is predictable and controllable, but it breaks down in unusual scenarios, and players quickly sense the artificiality. PUBG Ally isn't bound by a script. The language model generates a response for the specific moment: being surrounded in a late zone with no ammo calls for one tactic, an open field early in the match calls for another. Key capabilities of the system: Understands free-form voice commands with no predefined syntax Makes

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…