Developer abhi245y has released PS5LM Plus, a native GPU application that enables jailbroken PlayStation 5 consoles to function as local AI servers. The project builds upon Mert Cobanov’s PS5LM, which previously relied on CPU processing and achieved a generation speed of only one token per minute for 27B parameter models.
By integrating the AMD RADV driver from the Mesa project and utilizing the Vulkan compute interface, the developer successfully offloaded tasks to the PS5 GPU. This optimization increased available system memory from 6GB to 11.4GB, allowing the console to run highly quantized models like Qwen3.8-27B at a speed of 20.9 tokens per second.
The PS5 hardware demonstrates significant performance for smaller models, reaching 61 tokens per second for Gemma 4 E4B and 124 tokens per second for Llama 3.2 3B. During these operations, GPU utilization reaches 100 percent, while CPU usage remains at 2 percent, supported by the console's 448GB/s memory bandwidth.
PS5LM Plus includes a web-based chat interface and provides API services via port 8081 on a local network, allowing external devices to interact with the models. It also supports compatibility with OpenAI and Anthropic APIs, enabling integration with tools like Claude Code and Codex.
Despite these capabilities, the project is limited by the console's 16GB of unified RAM, which restricts model selection and quality. Furthermore, the requirement to jailbreak the console to run native llama.cpp introduces risks, including potential hardware bans and the suspension of PlayStation Network accounts.
