Key information
Developer abhi245y has introduced PS5LM Plus, a native GPU application built upon Mert Cobanov’s PS5LM project. This modification allows jailbroken PlayStation 5 consoles to function as local servers for large language models, including the Qwen3.8-27B model.
The project utilizes the RADV open-source AMD driver and the Vulkan interface to enable native llama.cpp execution. By optimizing memory management, the developer increased available memory from 6GB to 11.4GB, facilitating the operation of highly quantized models.
Performance varies by model size; the Qwen3.8-27B-UD-IQ2_XXS version achieves approximately 20.9 tokens per second, while smaller models like Llama 3.2 3B reach 124 tokens per second. During operation, the console maintains a power draw of 86W and temperatures around 54°C.
PS5LM Plus provides a web interface and API support, enabling integration with tools like Codex and Claude Code. However, the process requires a jailbroken console, which carries the risk of PSN account suspension, and the 16GB memory limit restricts users to heavily compressed models.
