Studio Atelico opens its on-device AI engine to outside game developers

Studio Atelico has opened Early Access to its Atelico AI Engine, making an on-device generative-AI toolkit available to outside game developers using Unity, Unreal Engine and Godot. The software runs AI workloads locally inside the game process, rather than relying on cloud data centres, and is intended for Windows, macOS, iOS and Linux across CPU and GPU hardware.

The engine first powered Atelico’s unreleased game “Bobium Brawlers”, according to MCV/DEVELOP, which reported the Early Access launch on Oct. 29. In that game, players describe a creature in natural language and receive generated character artwork and a playable deck on their phones.

Atelico is pitching the engine as infrastructure for AI features that work alongside a game’s rendering loop. Its feature set includes language models, image generation, speech recognition, text-to-speech, semantic search, facial animation and an NPC Character Behavior System. Developers can define characters’ personas, motivations, relationships, emotional states and memories, while contextual retrieval can bring authored dialogue and lore into generated responses.

The company also offers structured outputs for game systems and configurable content safeguards. The listed controls include keyword blocklists, machine-learning classifiers, LLM judges and content rewriters, with presets called ChildSafe and GameSafe. Its Atelico IG image model is advertised as generating images in about a second on local hardware, for uses including textures, item icons, character portraits and loading screens.

The central technical claim is that Atelico’s proprietary Compute Graphics Scheduler, or CGS, arbitrates between rendering and inference on a single shared hardware context. The company says conventional tools can trigger operating-system-level context switching and compete for GPU resources, while its system runs in-process alongside the renderer.

Atelico’s own RTX 5080 demonstration illustrates both the claimed benefit and the limits of the evidence. In a Windows 11 D3D12 test using Llama 3.2 3B and a San Miguel scene, it reported 123 frames per second without inference. Running AI without CGS cut the render loop to 3 FPS, while CGS in its balanced mode held the scene at 121 FPS as the model generated text. The results remain company benchmarks, however, and have not established performance across a broad range of consumer hardware and games.

The company also reports faster decoding in 22 of 24 tests across 12 models on CUDA and Apple Silicon, with ties in the other two. In one M1 Max example, prefix caching cut a roughly 1,000-token persona prompt’s time to first token from 347 milliseconds to 83 milliseconds. These are again Atelico’s reported results rather than independent performance tests.

For developers, the local approach is meant to remove network latency and per-request inference charges while keeping player data on the device. The engine supports native integrations through Godot 4’s GDExtension, Unity C# P/Invoke bindings and Unreal Engine 5 subsystem and Blueprint-compatible tools. Developers can request Early Access through Atelico’s website, but the studio has not announced a broad commercial release or full licensing terms.

Atelico has also tied its own game development to a draft Artist Rights Contract, or ARC. The proposed agreement sets out explicit consent, compensation and project-specific restrictions for artists whose work is used in AI development; it is not a general policy governing all uses of the engine. ARC currently applies only to “Bobium Brawlers”, and other uses of an artist’s work require renegotiation.

The company says it trains its own models on its own art and data, and says the aim is not to replace people. Piero Molino, Studio Atelico’s chief executive, said studios’ success would depend on using AI “entirely in service of the player experience,” rather than simply using the most AI. Alongside “Bobium Brawlers”, Atelico says its technology is running in GARP, a life-simulation demonstration with hundreds of agents and emergent dialogue rather than scripted beats.