Last year, our studio showcased a tech demo of generative agents running on-device. Fast forward a year later, we launched our first-ever game using this same approach. Bobium Brawlers is a mobile game with generative AI running fully on the player’s phone: no cloud, no data center. This talk is about how we made it, what realtime GenAI enables, and how we solved the technical challenge of making models small enough to fit in a phone to generate high quality outputs. We want to share the key decisions we made (and the tradeoffs considered) during development, and share our experience for any studio thinking about putting GenAI models on-device.
1. How to draw the trust boundary: a repeatable way to decide what a generative model should own versus what you should author or script, and why we think getting this wrong is likely the most common failure. 2. How to design for non-determinism so a model’s mistakes still feel intentional, using guardrails and fallbacks rather than hoping the model behaves. 3. The real on-device performance and resource budget on a phone and concrete ways those constraints should shape your game loop before you write a line of code. 4. How to beat the novelty cliff: turning one-time generative “wow” moments into a loop with replay value. 5. What going fully on-device actually buys a studio, and what it costs, stated as lessons learned vs a sales pitch.
