The author forked the PS5LM llama.cpp port to run Qwen 27B as a native PS5 application, leveraging the GPU via a Vulkan RADV port instead of the CPU‑only payload. This change increases available memory from about 6 GiB to roughly 11.4 GiB, enabling larger models. The resulting setup achieves inference speeds of around 20 tokens per second on a jailbroken PlayStation 5.