Intelligent NPCs
Atelico AI Engine: on-device NPCs with memory
Studio Atelico has opened Early Access to an engine that runs language models, speech, images and NPC behaviour on the player's own device, with plug-ins for Unity, Unreal and Godot.
On 29 October Studio Atelico opened Early Access to the Atelico AI Engine, a toolkit for putting language models, speech, image generation and character behaviour inside a game, running on the player’s device rather than in the cloud.
Access is by request. Atelico says it is a small team and places are limited, so this is not yet a public download.
How it works
The engine runs in-process, next to the renderer, with native integration for Unity, Unreal Engine and Godot. It targets Windows, macOS, iOS and Linux, on GPU or CPU. Because inference is local, player data stays on the device and there is no per-token bill.
The package covers dialogue models with structured output, image generation with custom style adapters, text-to-speech, speech-to-text, audio-driven lip-sync, and embeddings for search and memory across text, images and sound. Guardrails check the player’s input and the model’s output against the developer’s content rules.
The part most relevant to NPCs is the Character Behavior System. It gives characters personas, needs, emotions, relationships and memories that shape what they do next. Atelico’s example is a smuggler under questioning. The engine looks up the lines the writers authored for the scene and either returns one verbatim, when it matches the player’s question by meaning, or uses them as style examples for generation. The smuggler moves from calm to cornered, remembers an earlier betrayal, and has any off-script answer rewritten in character by the guardrails.
A second piece, the Compute Graphics Scheduler, lets the AI share the GPU with the renderer instead of taking turns with it.
The numbers
All figures are Atelico’s own.
- On an RTX 5080 in a 4K scene with 10 million triangles, running a language model as a separate program dropped the frame rate from 123 FPS to 3. With a 3B model and the scheduler set to Balance, the same scene held 121 FPS.
- Across 12 models on CUDA and Apple Silicon, Atelico says its engine decodes faster on 22 of 24 runs and ties on the other two.
- An image takes about 1 second on an M1 Max, about 4 seconds on an iPhone 13 and about 0.27 seconds on an RTX 5080.
- Speaking one second of dialogue takes about 56 ms, and the smallest language model is about 250 MB.
Limits
Early Access is gated, and the benchmarks are the vendor’s, run on its own workloads. The company has not named a third-party game using the engine; it says other developers are building with it but it is too early to show them. It runs in the studio’s own games, Bobium Brawlers and the life-sim GARP.
Small on-device models also trade away some capability. A 250 MB model will not write like a large cloud model, which is why Atelico leans on authored lines and guardrails.
Our take
The useful idea is the split between authored and generated content. Writers supply the lines and the rules, and the model chooses and shapes them. That is a practical answer to the question of how an NPC stays in character over a long game, and it connects to the problem we describe in what makes an NPC intelligent. Local inference also changes the economics, since a game that never pays per token can let characters talk freely. The open question is how much memory such characters can carry across a whole playthrough on a phone-sized budget, which we cover in memory in AI games.