On June 17, 2026, KRAFTON opened a two-week global beta of Ally Duo, a new Arcade mode for PUBG: BATTLEGROUNDS. In Ally Duo, your squadmate is an AI character named Ella who follows spoken instructions. This article compares KRAFTON's two releases with Ubisoft's NEO NPC prototype, and it explains in plain terms where each model runs: on your own graphics card or on a remote server.
Ella takes voice orders, and the model runs on your machine
KRAFTON says the AI in Ally Duo is powered by an on-device small language model built on NVIDIA ACE. In plain terms, the model runs on the player's own hardware. It doesn't send each request to a remote server. You team up with Ella on the Sanhok map and talk to her about movement, item collection, and strategy.
The design splits the work by how fast a response has to be. Ella uses speech-to-text and text-to-speech to talk with you. She is built to react with minimal latency for split-second actions like grabbing loot. For strategy, she synthesizes in-game context and conversational flow to make decisions. That split is a sensible way to handle a slow language model in a fast shooter.
The two companies don't use the same name for this kind of character. NVIDIA's ACE page calls them Co-Player Characters, AI-driven allies that communicate in natural language, while KRAFTON calls PUBG Ally a Co-Playable Character. It's the same idea under two labels.
Ally Duo was also a limited test. It supported Korean, Chinese, and English and ran from June 17 to July 1. KRAFTON's Chief AI Officer Kangwook Lee called it "an early step toward exploring how AI companions can create new ways to play."
Smart Zois: AI that plans your neighbors' day
inZOI went first. In March 2025, NVIDIA described KRAFTON's life sim as shipping the first NVIDIA ACE autonomous game characters. In inZOI, players turn ordinary townsfolk into AI-driven characters by activating a Smart Zoi setting.
What these characters do is less dramatic than Ella, and arguably more useful in a life sim. You give a Smart Zoi life goals through a text prompt, and a language model turns those goals into actions. NVIDIA's example: a Zoi told to do acts of kindness might offer food to a hungry stranger. At the end of each in-game day, the model helps adjust the Zoi's schedule of activities based on what happened that day. NVIDIA's ACE page sums up Smart Zois as agents that plan, act, and reflect on their decisions. The model runs on your own machine: NVIDIA says a half-billion-parameter small language model, Mistral-Nemo-Minitron Instruct, powers Smart Zois, and that Smart Zoi is GPU-accelerated exclusively on NVIDIA RTX GPUs. So you need an NVIDIA RTX card to use it.
So you don't have conversations with Smart Zois. The AI shapes how they behave in the simulation, and that's a much easier problem to keep under control than open-ended dialogue.
NEO NPC: strong conversations, no release
Ubisoft's project aims at the hardest version of the problem: free conversation. A small R&D team at Ubisoft's Paris studio built NEO NPC with Nvidia's Audio2Face application and Inworld's Large Language Model. The goal was to let players talk to characters without breaking the authenticity of the situation.
Most of the work went into guardrails. Data scientist Mélanie Lopez Malet trains the model to stay in character through guardrail systems, player input analysis, 3D environment and textual instructions. Filters catch toxic input, and when a player turns abusive, the character will no longer be collaborative. Mosser, who wrote the original character, put the limits plainly: "these characters do not have free will."
Ubisoft states the project's status itself: NEO NPC is only a prototype, with "still a way to go" before it appears in a game. Production technology SVP Guillemette Picard said you don't need to be a AAA developer to onboard this technology. That makes it a goal for indie developers, not something they can use today.
| Dimension | PUBG Ally (Ella) | inZOI Smart Zois | Ubisoft NEO NPC |
|---|---|---|---|
| Status | Global beta, June 17–July 1 | Shipped in-game setting | Prototype, no release |
| Character role | Squad teammate | Autonomous townsfolk | Story conversation partner |
| How you interact | Voice commands | Text-prompted life goals | Free conversation |
| Where the model runs | On-device small model | On the player's NVIDIA RTX GPU | Not stated by Ubisoft |
| Named tech | NVIDIA ACE | NVIDIA ACE | Inworld LLM, Audio2Face |
Does the AI run on my GPU or in the cloud?
It depends on the studio, not the toolkit. NVIDIA describes ACE as offering ready-to-integrate cloud and on-device AI models for speech, intelligence, and animation. NVIDIA says its on-device models are tuned for gaming hardware, with a small memory footprint.
The tricky part is that your GPU is already busy drawing frames. NVIDIA's In-Game Inferencing (NVIGI) plugins schedule AI inference alongside graphics work so the two don't compete for the card. The Total War: PHARAOH advisor shows the on-device approach working: it runs on a small on-device language model connected to Total War's extensive game data. NVIDIA's vision model, Nemovision-4B Instruct, can be deployed for cloud or on-device through the same SDK.
For indie developers, the tools are real but low-level. The Game Agent SDK (Beta) is a lightweight C/C++ agentic framework with Agent, Chat and RAG APIs. The In-Game Inferencing SDK runs models locally via in-process (C++) execution and CUDA in Graphics. CUDA is NVIDIA's technology, so take that into account when deciding which hardware your players need.
The tradeoff, roughly: running the model on the player's device uses graphics memory and frame time but avoids server bills and network lag. Running it in the cloud reverses that.
Which one matters to you
If you're a player who wants to try AI characters now, inZOI's Smart Zoi setting is the one available as a feature inside a shipped game. Ally Duo shows what an AI teammate in a shooter feels like, but it was a two-week beta, so keep an eye on KRAFTON for what comes next. NVIDIA also lists Wemade Next's MIR5, whose AI-powered bosses learn from previous player tactics. It also lists Dead Meat, an interrogation game where you can ask the suspect absolutely anything.
If you're an indie developer, follow KRAFTON's model, not Ubisoft's. A small on-device model, a narrow role, and fixed limits on what the character can do are achievable with NVIDIA's SDKs today. Treat NEO NPC as a guide to designing guardrails: it shows the limits you'd need around a character that talks back. Just don't plan around it as a tool you can get.
If you're waiting for fully open-ended story NPCs, you're still waiting. The AI characters that have actually shipped work because their jobs are narrow.
A gaming PC with its side panel removed, showing a large graphics card lit inside the case, with a dim server aisle in the background.
Sources
- developer.nvidia.com – "NVIDIA ACE for Games"
- news.ubisoft.com – "How Ubisoft’s New Generative AI Prototype Changes the Narrative for NPCs"
- krafton.com – "KRAFTON INTRODUCES PUBG ALLY BETA TEST | KRAFTON" (2026)
- developer.nvidia.com – "NVIDIA RTX Advances with Neural Rendering and Digital Human Technologies at GDC 2025 | NVIDIA Technical Blog" (2025)