Laya 421M Model Plays Flappy Bird Locally on CPU via OpenVINO INT8
A developer demonstrated running the 421-million-parameter Laya decision model quantized to 8-bit integers (INT8) using Intel's OpenVINO toolkit to play Flappy Bird in real time on a 12th Gen Intel Core i7 desktop CPU. This demo highlights how small, decision-focused AI models combined with CPU-optimized INT8 quantization can execute real-time autonomous agent tasks on standard consumer desktop hardware without needing a dedicated GPU. It shows the growing viability of lightweight local AI edge deployments for low-latency interactive applications. The project involved converting the Laya System One checkpoint into OpenVINO format and applying post-training INT8 quantization to dramatically reduce inference latency and memory overhead. Running directly on a 12th Generation Intel Core i7 desktop CPU, the setup achieved fast enough forward passes to make frame-by-frame game decisions.
## BACKGROUND
Laya is an open-source 421M-parameter decision model optimized for rapid local routing, scoring, and classification tasks rather than open-ended text chat. Intel OpenVINO is an open-source toolkit designed to optimize and accelerate deep learning model inference across Intel hardware architectures. INT8 quantization converts network parameters from 32-bit floating-point precision to 8-bit integers, greatly cutting memory footprint and compute demands with minimal loss in model quality.