IT之家 AI

Developer turns a hacked Sony PS5 into a local AI server: uses GPU acceleration to run the 27B-parameter Qwen3.8-27B model at 20.9 token/s

IT Home reported on October 17 that, according to foreign media Wccftech, developer abhi245y has developed a native GPU application, PS5LM Plus, based on PS5LM (a llama.cpp port project) by senior AI engineer Mert…

Image source · IT之家 AI

IT Home On October 17, according to foreign media Wccftech, developer abhi245y developed the native GPU application PS5LM Plus based on PS5LM (a llama.cpp port project) by senior AI engineer Mert Cobanov, allowing hacked PS5 consoles to run large language models such as Qwen3.8-27B locally.

IT Home learned that early versions of PS5LM mainly relied on the CPU to run models; when executing a 27B-parameter model, it could only generate about 1 token per minute on average. To improve performance, developer abhi245y ported the AMD RADV driver from the Mesa open-source project and used the Vulkan compute interface to let llama.cpp run as a native application on the PS5. This change increased available memory from about 6GB to 11.4GB, enabling the console to run highly quantized Qwen3.8-27B models.

Thanks to the PS5's high memory bandwidth (448GB/s), Qwen3.8-27B can generate at up to 20.9 tokens per second. During testing, GPU usage reached 100% while CPU usage was only 2%. For smaller models, the PS5 can achieve even higher generation speeds, for example Gemma 4 E4B at about 61 tokens per second and Llama 3.2 3B reaching 124 tokens per second.

Besides using the PS5's GPU to accelerate model inference, PS5LM Plus also includes a built-in web chat interface and can provide API services via port 8081 on the local network; users can access the page from other devices on the same network and chat with the locally running model. The project is also compatible with the OpenAI and Anthropic APIs, making it convenient for tools such as opencode, Codex, and Claude Code to call models running on the PS5.

For a console primarily designed for gaming, the PS5 being able to serve as a local AI server, providing GPU acceleration for large model inference, demonstrates its hardware's application potential beyond gaming. However, the console's 16GB of unified RAM still limits model choices and running quality. In addition, running the native version of llama.cpp requires jailbreaking the PS5 first, which may lead to risks such as the console and PSN account being banned.

Original source

IT之家 AI

Content notes

Original publication and rights belong to the source.

Machine translation · Refer to the original