Developer Uses Hacked PS5 as Local AI Server

A developer, abhi245y, has adapted a modified PlayStation 5 console to operate as a local AI server. The project, named PS5LM Plus, enables the execution of large language models, such as the 27-billion parameter Qwen3.8-27B, directly on the console's GPU.

Earlier iterations of the PS5LM project primarily relied on the CPU, achieving speeds of approximately one token per minute for the 27B parameter model. This enhancement utilizes AMD's open-source drivers and the Vulkan compute interface, increasing usable memory from around 6GB to 11.4GB. This allows for the operation of highly quantized versions of large models.

Benefiting from the PS5's substantial memory bandwidth (448 GB/s), the Qwen3.8-27B model can achieve generation speeds of 20.9 tokens per second. During testing, GPU utilization reached 100%, while CPU usage remained at only 2%. Smaller models, like Gemma 4 E4B, can attain speeds of up to 61 tokens per second.

PS5LM Plus also includes a web-based chat interface and an API service accessible via LAN port 8081. It supports OpenAI and Anthropic API formats, facilitating integration with various tools. While this demonstrates the PS5's hardware potential beyond gaming, the console's 16GB of unified memory still imposes limitations on model selection and quality. Modifying the console requires jailbreaking, which carries risks of console or PSN account bans.