PS5LM enables jailbroken PlayStation 5 consoles to run llama.cpp language models natively on the console's GPU, serving AI models to devices on the same network. The system leverages the PS5's RDNA 2 GPU and 11.4 GB of shared memory to load GGUF models and respond to requests without needing a PC, cloud, or account, according to github.com.
The software uses llama.cpp's Vulkan backend to operate directly on the PS5 GPU, supporting all architectures and quantizations compatible with llama.cpp. It provides API compatibility with OpenAI's /v1/chat/completions and Anthropic's /v1/messages endpoints, allowing any device on the network—laptops, phones, or coding tools—to interact with the model via a browser or API client pointed at the console's IP and port 8081.
PS5LM also supports decision models powered by Ollaya, including Jev-like decision models such as d1, Clef, and OpenJev. These models run on the GPU and answer typed decisions with calibrated probabilities rapidly, using the /v1/systemone API. This functionality extends the console's AI capabilities beyond chat to include yes-or-no and scoring questions about text and images.
The project demonstrates how a consumer gaming console can be repurposed as a private AI model server, eliminating reliance on cloud infrastructure. The latest update on github.com shows PS5LM running Qwen 3.5 9B models on a PS5 Slim, highlighting the system's versatility and performance.