Bijan Bowen first tests a compact quantized local model on a racing game built with Blender and Godot. It produces a basic playable result but needs guidance from a stronger model, illustrating how local inference can offload some work without replacing current frontier systems.
A larger mixture-of-experts model running through Strata produces a more detailed game and iteratively repairs selection-screen and wheel problems. Bijan Bowen reports roughly 160 to 170 tokens per second on a system with 32 GB of GPU memory and 64 GB of system RAM, while showing remaining steering, map and gameplay defects.
Bijan Bowen then repeats the task with Opus 4.6 at medium reasoning effort and encounters launch and collision problems. He presents the comparison as one exploratory test against an older frontier model, not a comprehensive benchmark or proof that local models match today's strongest hosted systems.
Watch on YouTube
