PS5LM enables native GGUF model execution on jailbroken PS5

PS5LM is a new application that allows users with a jailbroken PlayStation 5 to run large language models (LLMs) natively on the console's hardware. By implementing llama.cpp on the PS5, the project transforms the gaming console into a private model server. Once configured, the PS5 can serve AI responses to laptops, phones, and coding tools across a local network without requiring a PC, cloud connection, or account.

The application supports GGUF models and provides a TV-optimized interface. This interface allows users to monitor the system in real-time, download models directly from Hugging Face, and interact with the AI using a controller. The project also introduces support for decision models via Ollaya, which provide typed decisions and calibrated probabilities instead of standard text generation.

PS5LM utilizes the llama.cpp Vulkan backend to run models on the PS5's RDNA 2 GPU. The system allocates 11.4 GB of the console's shared memory for the model and its context. Because it uses llama.cpp, the application supports any architecture and quantization method known to that framework.

To make the console useful as a server, PS5LM implements several API endpoints. It supports OpenAI's /v1/chat/completions and Anthropic's /v1/messages, making it compatible with tools like Claude Code. Users point their clients to the console's IP address on port 8081.

For decision models, the app uses the /v1/systemone API. These models, such as d1 Omni 600M, process a state and answer specific questions in a single forward pass. This allows for fast, non-generative outputs, such as determining if a customer is angry or identifying a specific intent from a text string.

Performance varies by model size. On a PS5 Slim, the Qwen 3.5 0.8B model (Q4_K_M) reaches 207 tokens per second. The larger Qwen 3.5 9B model (Q4_K_M) runs at 55 tokens per second. For very large models, the Qwen 3.8 27B (UD-Q2_K_XL) operates at approximately 20 tokens per second with a 64k context window. Decision models are even faster; the d1 Omni 600M can answer three questions in 0.3 seconds.

PS5LM is not available for standard retail consoles. It requires a PS5 running firmware versions between 7.00 and 13.60 that has been jailbroken. The installation process is complex and requires several specific payloads to be loaded into the system memory.

Required payloads include kstuff-lite for kernel patches, etaHEN for homebrew support, and ShadowMountPlus to place the application on the console's home screen. Additionally, the developer recommends using elfldr on port 9021 to monitor temperature and power, and ftpsrv on port 2121 to transfer the application and model files from a PC to the console's internal storage, M.2 drive, or a USB drive.

Users must also manage memory constraints. While the system provides 11.4 GB of shared memory, the application's planner must calculate the longest possible context and best cache that can fit within the available free memory for each specific GGUF model loaded.

PS5LM turns a jailbroken PS5 into a capable local AI server by leveraging the RDNA 2 GPU and llama.cpp.