A jailbroken PlayStation 5 can now execute large language models directly on its hardware, using the open-source llama.cpp library to leverage the console's RDNA 2 GPU. The implementation, called PS5LM, loads GGUF-format models onto the console's 11.4 GB of shared memory and serves them across a local network without requiring external servers or cloud services.

The system enables private, on-premises AI processing where users can run models ranging from small 800-million-parameter variants to 27-billion-parameter configurations. Performance varies by model size: a 9-billion-parameter Qwen 3.5 model processes at approximately 55 tokens per second, while smaller 800-million-parameter versions achieve around 207 tokens per second. Decision-focused models like d1 Omni 600M complete three questions in 0.3 seconds, including image analysis tasks.

Technical implementation relies on llama.cpp's Vulkan backend to access the PS5's GPU capabilities. The software stack requires custom homebrew components including kstuff-lite kernel patches and etaHEN for homebrew execution. Once loaded, the system exposes standard AI APIs including OpenAI's /v1/chat/completions and Anthropic's /v1/messages endpoints, allowing compatibility with existing client applications.

Users interact through a TV-optimized interface that displays model activity, system metrics, and chat histories. The platform supports both language models and multimodal systems capable of processing images. Model selection occurs through an on-screen interface where compatible GGUF files are automatically downloaded from Hugging Face repositories.

Operation requires firmware version 7.00 through 13.60 on a jailbroken console. Setup involves copying PS5LM application files via FTP to the /data/homebrew directory, with additional components like ShadowMountPlus enabling home screen access. Updates are applied by replacing the application folder while preserving stored models and configurations.

The project maintains no affiliation with Sony Interactive Entertainment, though it utilizes unlicensed modifications to access the console's hardware features. All software components preserve their original licensing terms, with llama.cpp retaining its MIT license and other dependencies using compatible open-source agreements.