abhi245y/ps5lm-plus is a C++ project that runs llama.cpp as a native app on a jailbroken PlayStation 5, executing GGUF models on the console GPU. The latest release the repository points to is v0.8.3. It aims to run arbitrary GGUF models on the console, but what changed in this release is not stated in the provided information.
The core of this project is using the PS5's own GPU and 12 GiB of unified memory for inference rather than a PC. According to the description, earlier payloads could not use the GPU and had only about 6 GiB of memory, and this title app approach addresses that constraint. The implementation applies ggml's Vulkan backend to a RADV port and uses a single 12 GiB device heap. The app offers a TV dashboard, a model library, and download features, and manages GGUF models on internal storage, M.2, and USB drives. On networked devices, web chat and an OpenAI-compatible API are available. The license is GPL-3.0.
The model library computes and shows the suitable context length and KV cache type for each drive, and for HuggingFace downloads it reads the first 16 MiB to check the size before verifying with SHA-256. This is a design tailored to console hardware, but the performance figures rely on the README's own description. So it is limited as a basis for judging actual performance, and the specific changes in v0.8.3 could not be confirmed either.