NVIDIA ACE Explained (2026): What It Is, Which Games Use It, and What It Needs to Run
NVIDIA ACE is a set of AI models and developer tools for building game characters that listen to the player, reason, speak and animate their own faces. It is not a game. It is a toolkit that game studios build characters with — and since 2025, it has shipped mostly as small models that run on your own graphics card rather than in the cloud.
That second part is the important one, and it is not obvious from the name. ACE was announced in 2023 as the Avatar Cloud Engine. The characters actually shipping with it in 2026 run locally, on models small enough to share a gaming GPU with the game itself. And the best-documented of them, PUBG’s AI teammate, is built on a principle worth knowing whatever you play: the language model supplies the voice and the judgement, but the game engine keeps the facts.
This page is built from NVIDIA’s own announcements, developer pages and code, and from KRAFTON’s technical account of how PUBG Ally was built. We have not played the games below for review, and none of them is scored here.
NVIDIA, ACE, GeForce and RTX are trademarks of NVIDIA Corporation. Arcanum is independent and unaffiliated with NVIDIA and with every studio named here; names are used only to identify the products.
The Short Answers
- What is it? NVIDIA’s toolkit for AI game characters: speech recognition, a language model, text-to-speech and facial animation, plus the software to run them inside a game.
- What does ACE stand for? Avatar Cloud Engine, when it was announced in May 2023. NVIDIA’s current pages just call it ACE.
- Which games use it? PUBG (PUBG Ally), inZOI (Smart Zois), NARAKA: BLADEPOINT MOBILE PC VERSION, MIR5, Total War: PHARAOH and the upcoming Dead Meat, according to NVIDIA.
- Is it free? For players, it comes with the game. For developers, the Unreal Engine plugins are MIT-licensed and the Game Agent SDK is Apache 2.0.
- What do I need? A GeForce RTX card. PUBG Ally needs at least 8GB of video memory.
What ACE Was, and What It Became
2023: the Avatar Cloud Engine
NVIDIA announced NVIDIA Avatar Cloud Engine (ACE) for Games at Computex on 28 May 2023, describing it as “a custom AI model foundry service that transforms games by bringing intelligence to non-playable characters.” It bundled three existing NVIDIA products:
- NeMo, for building and customising language models, which “can be customized with lore and character backstories”;
- Riva, for speech recognition and text-to-speech;
- Audio2Face, for generating facial animation from a voice track.
The launch demo was Kairos, built with the startup Convai: a ramen shop in Unreal Engine 5 where the owner, Jin, “replies to natural language queries realistically and consistent with the narrative backstory.” The models could be deployed through NVIDIA’s DGX Cloud, on GeForce RTX PCs, or on premises.
2026: small models on your graphics card
NVIDIA’s developer page today describes ACE differently. It is “a suite of AI technologies designed to help middleware and game developers build knowledgeable, actionable, and conversational in-game characters,” providing “ready-to-integrate cloud and on-device AI models for speech, intelligence, and animation.” But the emphasis has moved: the first benefits it lists are “small models for gaming”, “optimized for on-device inference”, and running AI “alongside graphics” on the same card.
The shipped characters show how small. According to NVIDIA and KRAFTON:
| Character | Game | Language model | Where it runs |
|---|---|---|---|
| Smart Zoi | inZOI | Mistral NeMo Minitron, 0.5 billion parameters | Your RTX GPU |
| PUBG Ally | PUBG: Battlegrounds | Mistral-NeMo-Minitron-2B, 2 billion parameters, quantised | Your RTX GPU |
| SDK default | NVIDIA Game Agent SDK | Qwen3.5-4B, 4-bit quantised, about 3.8 GB of video memory | Your RTX GPU |
These models are far smaller than the ones behind the chat assistants most people know, and that is deliberate. KRAFTON explains why in its account of building PUBG Ally: cloud models were too slow. “The combination of network latency and model inference latency often made responses feel too slow for live squad communication.” Running a small model on your own card removes the round trip.
What ACE Is Made Of
In plain terms, an ACE character has four parts. A game can use any of them without the others.
- Ears — speech recognition. Turns what you say into text. NVIDIA ships an Unreal Engine 5 plugin “with a ready-to-use English language model.”
- Brain — a small language model. Decides what to say and do. The Unreal plugin provides “local, low-latency text generation and function calling.”
- Voice — text-to-speech. Speaks the reply, with ready-made voices.
- Face — Audio2Face-3D. Animates the character’s face to match the voice, including for Unreal’s MetaHuman characters.
Around those sit two pieces of engineering: the In-Game Inferencing SDK, which schedules the AI’s work alongside the game’s graphics on the same GPU, and the Game Agent SDK, a C/C++ framework whose APIs “help developers connect NPCs to game state, contextual knowledge and model-driven actions.”
That last phrase — connect NPCs to game state — is the part that decides whether an AI character is any good, and it’s where the best ACE game gets it right.
Which Games Use NVIDIA ACE
NVIDIA’s developer page lists its partner experiences by the role the AI plays:
| Role | Game | What the AI does, in NVIDIA’s words | Status |
|---|---|---|---|
| Teammate | PUBG: Battlegrounds | Co-Player Characters, “AI-driven allies that communicate in natural language and act autonomously like human teammates” | PUBG Ally beta ran 17–30 June 2026 |
| Citizens | inZOI | Smart Zois, “AI-driven agents that plan, act, and reflect on their decisions” | Optional experimental setting since March 2025 |
| Teammate | NARAKA: BLADEPOINT MOBILE PC VERSION | ”An on-device, NVIDIA ACE-powered AI teammate” | Added March 2025 |
| Enemies | MIR5 (Wemade Next) | “AI-powered bosses that continuously learn from previous player tactics” | Listed by NVIDIA |
| Assistant | Total War: PHARAOH | An AI advisor “powered by a small on-device language model connected to Total War’s extensive game data” | Listed by NVIDIA |
| Characters | Dead Meat | A murder-mystery interrogation game “where players can ask the suspect absolutely anything” | Announced for 2026 |
NVIDIA also lists a non-game partner: a streaming assistant from Streamlabs and Inworld AI.
One caution about “ACE games” lists. When ACE launched, NVIDIA named studios adopting Audio2Face — including GSC Game World for S.T.A.L.K.E.R. 2 and Fallen Leaf for Fort Solis. Audio2Face animates faces from recorded audio. It does not make characters you can talk to. A game can use one part of ACE and have entirely scripted characters, so check which part a game uses before expecting a conversation.
If what you actually want is a list of games where you can talk to the characters — ACE or otherwise — see our games with AI NPCs you can talk to.
How PUBG Ally Works — and Why It Can’t Be Talked Out of the Rules
PUBG Ally is the best-documented ACE character, because KRAFTON explained its architecture in detail to NVIDIA’s developer blog in June 2026. Three design choices stand out.
1. It splits fast reflexes from slow thinking. KRAFTON describes it as a “System 1 / System 2 problem.” A behavior tree — the traditional game-AI technique — handles “movement, aiming, and immediate combat responses at game tick rate.” The language model handles deliberate decisions and conversation. Nothing that needs to happen instantly waits on the model.
2. It only trusts the game engine. This is the key sentence in the whole interview: “The system prompt establishes that these tool results are the only ground truth for the current match.” Ally has tools to observe its own status, its teammate, nearby items and the fight, each returning engine data such as “You are armed with M416, 24/30 bullets in the magazine. HP 78%, inside the safe zone.” It is trained to re-check through those tools “rather than recalling from earlier turns.” So, in KRAFTON’s words, “Ally’s factual claims trace back to a value it deliberately pulled from the engine moments earlier, not to something the model guessed.”
3. It works in a closed world. Rather than handle every possible situation, KRAFTON “constrained the world: a single map, Sanhok, and a single mode, AI Duo, with a fixed item taxonomy.”
Compare that with the AI characters in Where Winds Meet, which players could complete quests with by simply describing actions they never took. Those characters appear to judge what you did from the conversation alone. PUBG Ally judges it from the engine. You can tell Ally you have a full magazine; the engine will tell it otherwise, and the engine wins.
That’s the same principle our benchmark rewards on its Determinism & Fairness axis, and the one independent research keeps arriving at: keep the game’s facts outside the model, and let the model narrate them. NVIDIA’s own SDK is built the same way — its sample assistant has a check_status tool for reading game state.
Memory, too. At CES 2026, KRAFTON showed Ally with long-term memory, and it is structured rather than a transcript: a player profile carried across matches (“name, preferred weapons, favorite drop locations”) and a record of past matches, plus short-term context within the current game. One playtester asked Ally to look out for a particular rifle in their first match; “from the next match onward Ally started finding it without being asked.” KRAFTON says it refined Ally with feedback from “over a thousand real players.”
What You Need to Run It
As a player:
- An NVIDIA GeForce RTX graphics card. ACE characters run on your own GPU. NVIDIA says PUBG Ally’s three models run “locally on RTX GPUs with at least 8GB of VRAM,” and that inZOI’s Smart Zoi feature is for “gamers with GeForce RTX GPUs.”
- Video memory to spare. The model shares the card with the game. KRAFTON notes that PUBG “already consumes a significant share of GPU memory,” which is why it quantised its model to fit “within the VRAM headroom that remains.”
- No subscription. Because the model runs locally, there is no separate AI service to pay for. It comes with the game.
Players on other hardware — AMD or Intel graphics, consoles, phones — are not the audience for the on-device features described here.
As a developer:
- Unreal Engine 5 plugins for speech recognition, the language model, text-to-speech and Audio2Face-3D, which NVIDIA describes as “local, RTX-optimized workflows bundled with ready-to-use models, governed by the MIT license.”
- The Game Agent SDK (beta), open source under Apache 2.0 on GitHub since May 2026. Per its documentation it needs Windows 10 or 11, “any NVIDIA Ampere GPU architecture (or newer),” driver 570.65 or later and CUDA 12.8. It bundles Qwen3.5-4B as its default model, plus retrieval models for grounding answers in a game’s own knowledge.
Where ACE Sits Among AI Character Tools
ACE is one of several ways to build AI characters, and the field has moved fast.
- Convai co-built the original Kairos demo in 2023 and still positions itself as “Conversational AI for Virtual Worlds.” Our Convai guide covers its plans and how its characters remember.
- Inworld AI, once best known for game characters, now sells “realtime TTS and STT models, LLM serving, and the inference behind both” for consumer applications — a broader business than games. It appears on NVIDIA’s page as the partner on the Streamlabs assistant. Our guide to Inworld AI dates the change from its own homepage.
- General-purpose language models reached through an API are what many smaller games use. Studios that tested cloud models for real-time play, like KRAFTON, found them too slow for combat.
Our research index tracks these alongside the academic work on AI characters, which reaches the same conclusions about keeping state outside the model.
If You Want AI Characters to Play With Now
ACE is mostly invisible to players: a teammate in one mode of PUBG, an experimental setting in inZOI. If you want more of this:
- Our list of games with AI NPCs you can talk to covers every game we could verify, ACE or not.
- If you want every character in a game to work like this, that is an AI RPG — our benchmark scores 27 of them on seven axes, from first-hand play.
Frequently Asked Questions
What is NVIDIA ACE? NVIDIA ACE is a set of AI models and developer tools for building game characters that can hear the player, reason, speak and animate their faces. NVIDIA describes it as a suite of AI technologies for building knowledgeable, actionable and conversational in-game characters, with cloud and on-device models for speech, intelligence and animation. It is a toolkit for game developers, not a game you can play.
What does NVIDIA ACE stand for? When NVIDIA announced it in May 2023, ACE stood for Avatar Cloud Engine, and it was presented as a custom AI model foundry service running in the cloud or on RTX PCs. NVIDIA’s current pages simply call it NVIDIA ACE, and the focus has moved to small models that run on the player’s own graphics card.
Which games use NVIDIA ACE? NVIDIA lists PUBG: Battlegrounds with PUBG Ally, an AI teammate; inZOI with Smart Zois; NARAKA: BLADEPOINT MOBILE PC VERSION with AI teammates; Wemade Next’s MIR5 with AI bosses; Total War: PHARAOH with an AI advisor; and the upcoming Dead Meat. Some other games have used only one part of ACE, Audio2Face facial animation, which is not the same as having characters you can talk to.
Is NVIDIA ACE free? For players, ACE features are part of the games that include them, and because the models run on your own graphics card there is no separate AI subscription. For developers, NVIDIA’s Unreal Engine 5 ACE plugins are released under the MIT license and the Game Agent SDK under the Apache 2.0 license, both of which allow free use.
What do I need to run NVIDIA ACE features? An NVIDIA GeForce RTX graphics card. NVIDIA says PUBG Ally runs locally on RTX GPUs with at least 8GB of video memory, and inZOI’s Smart Zoi feature is available to players with GeForce RTX GPUs. The AI model shares the card with the game, so it uses memory the game would otherwise have.
Can NVIDIA ACE characters be tricked like other AI NPCs? It depends on how the game uses them. PUBG Ally is built so that information pulled from the game engine is the only ground truth for the match, and a separate behavior tree carries out its actions. A character built that way cannot be talked into believing something that the game itself says is false, unlike AI characters that judge your actions only from what you tell them.
Sources
Checked 26 September 2026. NVIDIA: ACE for Games developer page; 2023 announcement (28 May 2023); inZOI and NARAKA launch post (13 March 2025); PUBG Ally beta post (17 June 2026); KRAFTON Q&A on building PUBG Ally (25 June 2026); Game Agent SDK (README and documentation). Convai and Inworld descriptions are from their own homepages.