AI teammates in games are moving from scripted followers toward characters that can interpret natural language, observe a changing world, and act without waiting for a rigid command menu. That shift matters on a platform with many playable worlds: an Elseland game session can begin with a genre choice, but memorable characters are often what make a player stay.
The clearest current example is PUBG Ally, a co-playable character developed by KRAFTON with NVIDIA ACE. Its June 17–30, 2026 public beta turned an industry demo into something players could test inside an actual high-pressure multiplayer game.
The important story is not that a bot can talk. It is that speech, game-state awareness, memory, navigation, looting, combat, and cooperation now have to work as one readable teammate.
Quick read
Key takeaways
- A co-playable character combines conversation with actions that affect the live game state.
- PUBG Ally uses a 2-billion-parameter on-device language model while a conventional behavior tree handles split-second movement, aiming, and combat reactions.
- A useful AI teammate must be responsive and grounded in the game before it becomes witty or broadly conversational.
- Players should be able to understand what the companion remembers, what it can do autonomously, and how to correct a bad decision.
From Scripted NPC to Co-Playable Character
Traditional companion NPCs usually select from authored dialogue and behavior branches. They can still be excellent characters, but their responses are bounded by states the design team anticipated in advance.
KRAFTON uses the term co-playable character, or CPC, for a different target. A CPC is expected to interpret the player’s intent, read current conditions, cooperate toward a goal, and choose useful actions. The character is not merely present in the scene; it participates in the loop.
| Character type | Understands | Acts through | Typical limitation |
|---|---|---|---|
| Scripted companion | Defined triggers and dialogue states | Authored behavior trees and animations | Cannot respond outside anticipated branches |
| Conversational NPC | Natural-language player input | Generated dialogue plus limited actions | May speak flexibly without helping gameplay |
| Co-playable character | Player intent plus live world state | Language reasoning, tools, and fast game AI | Harder to keep responsive, reliable, and predictable |
How the PUBG Ally AI Teammate Works
PUBG Ally accepts voice and text, understands PUBG terminology, map locations, item attributes, and live battlefield situations, and can make independent choices about looting, navigation, fighting, and support. The player does not need to prompt every step.
NVIDIA documents three local models in the June 2026 beta pipeline: speech-to-text, a 2-billion-parameter Mistral-NeMo-Minitron small language model, and KRAFTON’s text-to-speech model. The system targets RTX GPUs with at least 8 GB of VRAM.
The architecture deliberately keeps reflexes away from the language model. A behavior tree handles movement, precise aiming, and immediate combat reactions at game speed. The language layer interprets intent, plans deliberate actions, and generates speech. That division is a useful design lesson: not every intelligent-looking behavior should wait for generative AI.
| Pipeline layer | Input | Output | Why the player notices |
|---|---|---|---|
| Speech recognition | Player voice | Transcribed instruction | Commands feel natural only when transcription is reliable |
| Language reasoning | Instruction plus current match state | Plan, response, and tool calls | Advice must match what is actually happening |
| Behavior tree | Plan plus immediate game events | Movement, aiming, combat, and interaction | Fast actions cannot pause for a long model response |
| Text to speech | Generated response | Spoken teammate voice | Timing and tone determine whether dialogue helps or distracts |
The Five-Part Player Test for AI Companions
A demo can look impressive because the successful conversation was selected in advance. A player test is stricter: give the companion an ambiguous request, change the situation halfway through, and see whether it recovers without creating extra work.
Responsiveness is the first gate. NVIDIA’s technical Q&A says cloud-hosted language-model tests often felt too slow for live squad communication, while local inference produced more predictable response timing. For a battle royale, a shorter answer at the right moment can be more useful than a richer answer after the fight is over.
| Dimension | Player question | Healthy signal | Failure signal |
|---|---|---|---|
| Responsiveness | Does it answer and act in time? | Reaction fits the pace of the game | The moment has passed before the action starts |
| Grounding | Does it understand the current state? | Mentions real items, positions, threats, and goals | Invents objects or ignores visible danger |
| Autonomy | Can it help without micromanagement? | Completes a clear sub-goal independently | Waits for every tiny instruction or takes over completely |
| Memory | Does it learn useful preferences? | Recalls relevant playstyle or prior choices | Forgets immediately or stores unrelated details |
| Control | Can I redirect or stop it? | Correction is quick and understandable | Bad behavior repeats with no visible override |
Memory Should Improve Play, Not Collect Trivia
NVIDIA says PUBG Ally uses structured short- and long-term memory across sessions. The valuable memory is not a transcript of everything a player said. It is a small set of durable facts that improves future decisions: preferred pace, usual role, communication style, or a recurring strategic choice.
Roleplaying worlds create an even stronger case for memory because relationships, promises, rivalries, and quest decisions can shape later scenes. In AI RPG games, the best memory is visible through consequences rather than a character announcing that it has stored a fact.
A player should also know when memory is temporary, when it persists, and how to reset it. Long-term continuity feels personal only when it remains relevant and controllable.
Why Better Conversation Is Not Enough
A fluent character can still be a poor teammate. It may talk over combat audio, repeat an obvious plan, choose an unsafe route, consume a needed item, or generate a plausible explanation for an incorrect action.
Game designers therefore need hard boundaries around tool access, movement authority, resource use, friendly fire, story-critical decisions, and irreversible actions. Generative reasoning works best inside a clear ruleset with deterministic fallbacks.
The result may look less magical than an unrestricted chatbot, but it produces a more dependable game. Player trust comes from understanding what the companion can do and seeing it recover consistently when the unexpected happens.
- Keep reflex actions in fast deterministic systems.
- Ground decisions in a compact view of the real game state.
- Limit autonomous actions that consume rare resources or change permanent progress.
- Provide a fast stop, regroup, or follow command.
- Show uncertainty when the character lacks enough information.
What AI Teammates Could Change Next
The near-term opportunity is not a universal companion that can play every game. It is a domain-specific teammate that understands one world deeply, communicates at the right moments, and performs a bounded role well.
That pattern can extend beyond shooters: a navigator in an exploration game, a party member in an RPG, a coach in a strategy game, a resident in a life simulation, or a rival that studies the player’s repeated tactics.
For players, the standard should remain practical. Does the character make the next decision more interesting, create a story worth remembering, or let the player try a role that was previously inaccessible? If not, the AI layer is decoration rather than play.
Practical answers
Frequently asked questions
What is an AI teammate in a game?
An AI teammate is a computer-controlled character designed to cooperate with the player. Newer systems may combine conventional game AI with speech recognition, language models, memory, and tool use so the character can understand natural instructions and respond to live game events.
What is a co-playable character?
Co-playable character is KRAFTON’s term for an AI character that participates alongside a player rather than only delivering dialogue or following a fixed route. The goal is cooperative action, adaptive behavior, communication, and continuity across changing situations.
Is PUBG Ally a chatbot inside PUBG?
No. Conversation is one part of the system, but PUBG Ally also reads match state and can loot, navigate, fight, revive, and adjust strategy. A behavior tree still performs fast tactical actions while the language model handles intent and deliberate reasoning.
Does PUBG Ally run in the cloud?
The June 2026 implementation described by NVIDIA runs its speech, language, and voice models locally on supported RTX hardware. NVIDIA lists at least 8 GB of VRAM for that beta configuration.
Why use a small language model for a game companion?
A smaller local model can reduce network delay and use less memory than a large cloud model. In a real-time game, predictable responsiveness and knowledge of the specific game can matter more than broad general knowledge.
Can an AI teammate replace a human player?
It can fill a role, support practice, or make solo sessions feel more cooperative, but it does not reproduce the full creativity and social context of a human teammate. The better goal is a dependable companion for a defined situation, not a universal replacement.
What can go wrong with AI companions in games?
Common risks include slow responses, incorrect world knowledge, repetitive speech, unsafe autonomous actions, unwanted memory, and high hardware or service costs. Strong systems limit authority, expose correction controls, and fall back to conventional game logic.
Will every game use generative AI teammates?
No. Many games benefit from precise authored characters and deterministic behavior. Generative AI is most useful when interpreting flexible player intent or producing variation is essential to the core experience.
Checked July 28, 2026
Sources and further reading
- PUBG Ally Duo Mode beta
NVIDIA’s June 17, 2026 beta announcement and technical overview.
- How KRAFTON built PUBG Ally
Architecture, 2B-parameter model, memory, latency, and player-testing details.
- KRAFTON PUBG Ally playtest plan
Official CPC positioning, supported interactions, languages, and playtest plan.
Next step
