Llama.cpp on a jailbroken PS5: run any GGUF model natively on the console

The first release, on October 2, ran small Qwen models on the CPU only, with a 1.6 GB model cap. Version 0.9.0 moved inference to the console's RDNA 2 GPU through llama.cpp's Vulkan backend, with 11.4 GB of shared memory for the model and its context. Developer-reported figures on a PS5 Slim: about 207 tokens per second for a 0.8B model, 55 for a 9B model, and roughly 20 for a 27B model at 2-bit-class quantization with 64k context. The GPU work began in a fork by developer Abhiram, called PS5LM Plus. The console acts as a home server. Clients on the same network connect through OpenAI-compatible and Anthropic-compatible endpoints, or a browser chat page. Setup requires firmware 7.00 to 13.60 and homebrew payloads such as etaHEN and kstuff-lite. The speeds come from the developers, not independent benchmarks. The project is not affiliated with Sony. For an AI company, the product is not the PS5. The result shows that a consumer box with about 11 GB of usable memory can serve a 27B-class model at chat speed. Memory capacity, more than compute, sets the ceiling for cheap local inference. Standard API formats mean existing client code runs against such a device unchanged. The jailbreak requirement keeps this a hobbyist route, not a deployment target. Still, it is a reference point for how much model fits in console-class, fixed-memory hardware.