PS5LM Plus Lets Jailbroken PS5 Run 27B-Parameter LLM Locally

Developer abhi245y has released PS5LM Plus, a native GPU application built on Mert Cobanov's PS5LM project, allowing jailbroken PlayStation 5 consoles to run large language models such as Qwen3.8-27B locally.

The early version of PS5LM relied mainly on the CPU, generating roughly one token per minute when running a 27B-parameter model. To improve performance, abhi245y ported AMD's RADV driver from the Mesa open-source project and used the Vulkan compute interface to run llama.cpp natively on the PS5. This increased available memory from about 6GB to 11.4GB, enabling the console to run a heavily quantized Qwen3.8-27B model.

Thanks to the PS5's 448GB/s memory bandwidth, Qwen3.8-27B achieves a generation speed of 20.9 tokens per second. During testing, GPU utilization hit 100% while CPU usage was only 2%. Smaller models run even faster, with Gemma 4 E4B at about 61 tokens per second and Llama 3.2 3B at 124 tokens per second.

PS5LM Plus also includes a web chat interface and provides an API service on port 8081 over the local network, allowing other devices to interact with the model. The project supports OpenAI and Anthropic APIs, making it compatible with tools like opencode, Codex, and Claude Code. It is worth noting that running native llama.cpp requires jailbreaking the PS5, which may lead to bans on the console and PSN account.