AI NPC engines compared on the questions that matter
As of October 2026, you can put a language-model character in a game in five ways: a hosted character platform such as Inworld or Convai, an on-device stack such as NVIDIA ACE, a behavior engine such as Artificial Agency, a platform-native tool such as Epic’s Conversations in Fortnite, or your own code on a raw model API. None of them is the best choice for every game. Compare them on where inference runs, who keeps the persona and memory data, latency, the cost model, engine support, and what happens if the vendor stops.
Rick Vu, who directs this site, builds Red Dog Studio, a game studio building AI-native NPCs, so he has a stake in this market. For that reason this page does not rank anyone, and it says nothing about Red Dog’s own stack. Each claim below comes from the vendor’s own page or from a named report, linked in the sentence, and was read on 9 October 2026. Vendor pages change, so check them again before you commit.
A ranking would hide the fact that these tools answer different questions. A hosted platform sells a finished pipeline, an on-device stack sells a way to run models on hardware you do not own, and a raw API sells tokens and leaves the rest to you. Comparing them on demo quality alone tells you little about the game you will ship.
The five approaches in one paragraph each
A hosted character platform runs the model, the voice, and the character tooling on its own servers. Your game sends player input through an SDK and gets back a line and an action. Inworld now describes itself as “a research lab and inference provider focused on realtime AI models for consumer-facing applications,” with games as one use case among several. Convai calls itself a platform to “craft and deploy 3D AI characters” and lists long-term memory, a knowledge bank, and actions.
An on-device stack ships the model with the game. NVIDIA’s In-Game Inferencing SDK runs speech recognition, language, and embedding models inside a C++ game process.
A behavior engine gives characters goals and lets a model choose what they do. TechCrunch’s 2024 report on Artificial Agency says developers give each NPC “a set of motivations, rules and goals to live by.” Artificial Agency’s own news post on Owlchemy Labs says its engine will run the AI crew in a single-player mode of Dimensional Double Shift that is “slated to ship in the first half of 2027.”
A platform-native tool lives inside one game’s creator tools. Epic’s personas overview for Unreal Editor for Fortnite describes an LLM “that players can interface with in your island through a voice chat.”
A raw model API means your game or your server calls a language model directly. You write the prompt, the memory store, the checks, and the voice pipeline yourself. How AI NPCs work describes those parts.
Where inference runs
Inworld’s homepage calls its inference “Inworld-hosted” and lists deployment as “Cloud API + VPC (Enterprise).” Convai’s homepage lists “On Prem Deployment Options” among its enterprise features but does not say where the standard plans run. Both pages describe a service that your game reaches over the network.
NVIDIA’s stack is the on-device option in this list. The In-Game Inferencing SDK post says it “Supports both on-device inference (CPU/GPU) and cloud-based AI workflows.” Its model table lists Mistral-Nemo-Minitron at 2B, 4B, and 8B parameters, Llama-3.2-3b Instruct, and Whisper for speech recognition.
Epic’s documentation does not name the model or say where it runs. Outlook Respawn’s report on the Conversations launch says Epic uses Google’s Gemini 3.1 Flash-Lite for the character and ElevenLabs for the voice, which are both hosted services.
Artificial Agency splits the work. An AWS Startups article on Artificial Agency, dated May 2026, describes in-engine tools plus a cloud platform that manages agent state and memory. The same article says the company is fine-tuning Llama models at 1B, 3B, and 8B parameters and that the work was not yet in live deployment.
With a raw API, inference runs wherever your provider runs it. You can also host open-weight models on your own servers, which moves the bill from tokens to GPUs.
Who keeps the persona and memory
The persona is a block of text, and memory is a store of facts about the player. Where those live decides how hard it is to leave.
On a hosted platform, the character is defined in the vendor’s tools. Convai’s knowledge bank lets you “Upload all kinds of documents to add to character’s knowledge,” which means those documents sit with Convai. Artificial Agency’s cloud platform, per the AWS article, holds agent state and memory.
In Fortnite, the persona is a prompt inside your island, and Epic’s rules govern it. The personas overview says personas and prompts “must adhere to the Fortnite Developer Rules” and that “Starting in 41.30, Fortnite IP characters cannot be edited.”
On the device and on a raw API, the persona and the memory store are files and tables you own. You also own every bug in them.
Whichever route you take, keep the persona text, the memory schema, and the list of allowed actions in your own repository. They are the parts you would have to rebuild if you moved.
Latency
Every cloud route adds a network round trip to each turn. On-device inference removes that trip but competes with rendering for the same GPU. The In-Game Inferencing SDK post handles that with scheduling and a “vramBudgetMB” setting that caps the model’s memory.
Convai’s homepage promises “Low latency responses.” No vendor page read for this article publishes a latency figure you could compare with another vendor’s. Measure the time from the end of the player’s line to the first audible word in your own game, on your own minimum spec. AI NPC latency and voice covers where that time goes.
Cost model
The approaches bill in different units, so compare them on cost per player hour, not on the price list.
- Inworld’s pricing page sells monthly plans from $25 to $1,500 that include the same amount in usage credits. Text-to-speech bills per million characters, speech recognition per hour, and language models “at cost.”
- Convai’s homepage offers a free start and links to a pricing page. That page did not return readable plan details when fetched for this article, so check it directly.
- NVIDIA’s ACE developer page lists no price and says its Unreal Engine plugin suite is “governed by the MIT license,” while some of its models carry other licenses. The real cost is the player’s GPU memory and a higher minimum spec.
- Outlook Respawn reports that Epic pays the Gemini and ElevenLabs costs during the experimental phase.
- Artificial Agency publishes no price. In the 2024 TechCrunch report, an investor says studios will pay for licenses.
- A raw API bills per token. What an AI NPC costs per player works through the formula.
Engine support
Inworld’s games page says you can “Wire voiced characters into Unreal, Unity, a custom engine, or a media pipeline.” Its Character Studio deprecation page names a Unity AI Runtime SDK and an Unreal Agent Runtime SDK. Convai’s documentation lists Unity, Unreal Engine, and web plugins.
NVIDIA’s ACE page lists Unreal Engine plugins, including one for local text generation and function calling, and does not mention Unity. Artificial Agency’s March 2026 update announces a Unity plugin and says “outside Unity and Unreal, our API provides access to any game engine.” Epic’s Conversations work only in Unreal Editor for Fortnite.
None of the vendor pages read for this article names Godot. A Godot team today has two routes. It can call a model API over HTTP, or it can wrap a C++ SDK in its own engine extension.
What happens if the vendor stops
This question has a real example. Inworld’s Character Studio deprecation page says “The legacy Inworld Character Studio has been retired.” The page points developers to rebuild with Runtime SDKs and templates. It does not mention an export path for existing characters. That is an ordinary product change, but a game that kept its characters only in that tool would have to rebuild them.
The risk differs by route:
- On a hosted platform or a behavior engine with cloud state, NPCs stop talking when the service stops. Write fallback lines that fit the fiction for every character.
- On the device, the model weights ship in the build and keep running. You lose updates, not the feature.
- In Fortnite, the feature exists only as long as Epic keeps it, and Outlook Respawn reports that islands using it could not yet publish to public servers.
- On a raw API, the provider retires models over time. Because the prompt and memory are yours, moving to another provider is a code change and a round of testing.
How to choose without a ranking
Start from your game, not from the vendor list. If your players are on PCs with recent NVIDIA GPUs and the game must work offline, look at the on-device route first. If you want voice and character tooling working this month in Unity or Unreal, a hosted platform is the shorter path, and you should ask the vendor for an export format in writing. If you build in Fortnite, Conversations is the only route that fits inside the platform. If you need to control every part of the loop or you build in Godot, a raw API or an engine-agnostic API, such as the one Artificial Agency describes in its March 2026 update, is the likely route, and you write the parts a platform would have given you.
Run the same character through two routes before you decide. Use the same persona, the same ten player lines, and the same test machine. The comparison that matters is how that character behaves in your game.
Related reading
- How AI NPCs work, for the persona, memory, and checks that every route above has to provide.
- What an AI NPC costs per player, for turning a token price into a cost per player hour.
- AI NPC latency and voice, for where the delay in a spoken reply comes from.
Questions
What is the best AI NPC engine?
No single engine is best for every game. Decide first where the model should run, on a vendor's cloud, on your own servers, or on the player's device, because that choice settles latency, cost, and data ownership. Then pick among the tools that fit that answer and your game engine.
What are alternatives to Inworld?
As of October 2026 the main alternatives are Convai, a hosted platform with Unity, Unreal, and web plugins, NVIDIA ACE with its In-Game Inferencing SDK for on-device models, Artificial Agency's behavior engine, Epic's Conversations inside Fortnite, and your own code on a raw model API. Inworld itself retired its Character Studio and now offers Runtime SDKs and realtime voice APIs.
Can AI NPCs run without an internet connection?
Yes, if the model runs on the player's device. NVIDIA's In-Game Inferencing SDK runs language and speech recognition models inside the game process on the player's PC. Hosted platforms and model APIs need a connection for every turn.
Which engines support AI NPCs in Unity and Unreal?
Inworld and Convai both list Unity and Unreal support, and Artificial Agency announced a Unity plugin in March 2026 with an API for other engines. NVIDIA's ACE page lists Unreal Engine plugins and does not mention Unity. None of the vendor pages read for this article names Godot.