Atelico 发布 AI Engine 早期访问版

In Bobium Brawlers, players describe a creature in a few words and get a portrait and a playable deck to battle friends on their phone. To enable this without breaking the bank over cloud AI costs, we built the AI Engine that runs locally on the player device. We belive it can enable game developers to create entirely new types of gameplay like user-controlled procedural content generation like in Bobium Brawlers, but also smarter NPCs that remember interactions with the player and adapt dynamically their behavior, text and voice based interactions and a level of personalization that we have not seen in games yet. We are now opening Early Access to our AI Engine and can’t wait to see what other developers will build with it!

▶ Watch the engine trailer

Built especially for games

Cloud-based AI doesn’t play nicely with games: frame budgets are measured in milliseconds. The Atelico AI Engine runs in-process, right next to your renderer, with native integration for Godot, Unity and Unreal. It runs on Windows, macOS, iOS and Linux, on GPU or CPU. Player data never leaves the device, and you never pay per token. This frees game designers from cost considerations and provides them new tools for their creativity.

Out of the box, it brings everything a game needs to make AI part of its design: language models for dialogue and narrative, with structured output that always matches the shape your code expects, image generation with custom style adapters, text-to-speech and speech-to-text, lip-sync animation driven by audio, and embeddings for memory and search across text, images and sound. On top of those sit the pieces that make it shippable: guardrails that keep players’ input and the model’s output within your content rules, and a behavior engine that gives NPCs needs, emotions, memories and relationships.

AI that doesn’t steal your frames

The obvious way to add a language model to a game is to run it as a separate program next to the game. But the two then take turns on the GPU without coordinating, and while the model generates, your game waits. On an RTX 5080, in a 4K scene with 10 million triangles, that dropped the frame rate from 123 FPS to 3.

So we built the Compute Graphics Scheduler (CGS), which lets the AI share the GPU with your renderer instead of fighting it. With a 3B model generating text and CGS set to Balance, the same scene held 121 FPS.

And thanks to our megakernel compilation, it is extremely fast. Across 12 models on both CUDA and Apple Silicon, the engine decodes faster on 22 of 24 runs and ties on the other two. Gemma 4 E2B gets 41% faster on an RTX 3090. LLaMA 3.1 8B more than doubles on an M1 Max.

And the same applies to other modalities:

• ~1 second to generate an image on an M1 Max, ~4 s on an iPhone 13, ~0.27 s on an RTX 5080
• Up to 4× faster time-to-first-token in multi-turn dialogue, thanks to prefix caching
• ~56 ms to speak a second of dialogue, ~30 ms to transcribe one, under 3 ms per lip-sync frame
• ~250 MB for our smallest language model. Small enough to ship inside the game.

Your words, at the right moment

Performance is half the problem; the other half is getting AI to stay true to the game you’re crafting. The Character Behavior System gives NPCs personas, relationships, motivations and memories that shape what they do next.

Say your player is questioning a smuggler. The engine pulls the lines you wrote for this scene and either returns them verbatim when they match semantically with the player questioning, or are used as examples for informing the style of language generation. The smuggler goes from calm to cornered as the player pushes. He remembers the player double-crossed his brother two quests ago. If the player tries to steer him somewhere he’d never go, your guardrails rewrite his answer in character.

Production ready

The tech is already running in Bobium Brawlers and in GARP, our life-sim playground that runs hundreds of agents with emergent dialogue and no scripted beats. Other developers are already using our AI Engine for building their games (too early to show something, but stay tuned).

Come build with us

If you are developing a game that would benefit from our AI Engine with a novel game design, reach out to us! We’re a small team, so spots in the Early Access are limited, but we’ll read and respond to every request ourselves. Head to atelico.studio/ai-engine for code samples in Godot, Unity and Unreal, the full benchmarks, and to request access.

Got questions, or want to show us what you’re building? Come hang out on Discord.