Gpu

Steam Deck OLED tuning is no longer just about pushing sliders and hoping for more FPS. The stack is layered. Valve’s kernel, your Proton version, the game engine, and power policy all interact. Tune one layer alone and you often trade smoothness for crashes, or frame rate for battery drain.

This guide chases one goal: steadier frame times and longer battery life, without turning your Deck into a fragile science project. You get a safe workflow, specific kernel options, and game profiles you can reuse.

You can run Llama 4 Scout on a 24 GB consumer GPU, but only with an aggressive quantization and some patience. Scout is a 109B-parameter Mixture-of-Experts model, and even its smallest Unsloth dynamic GGUF build is about 32 GB, so a 24 GB card runs it with CPU offload at roughly 20 tokens per second. This guide covers which Llama 4 model fits your hardware, the real VRAM math, and the fastest way to get it running.