If you’re into mini PCs and local AI, you might be interested in a recent experiment where a jailbroken PS5 was turned into a local AI server capable of running a 27-billion-parameter language model (LLM). The PS5’s 16GB GDDR6 memory and 448GB/s bandwidth make it a decent candidate for this kind of task, but there are significant trade-offs.
The project, dubbed “PS5LM,” involved using the PS5’s GPU to handle the heavy lifting, freeing up memory for the AI model. Initially, the available memory was limited to just 6GB, but an optimized port increased this to 11.4GB. This allowed the PS5 to run a heavily quantized version of Qwen3.8-27B, which stores weights at roughly 2 bits each, shrinking the size to 6.8GB.
However, running a heavily quantized AI model comes with quality trade-offs, resulting in more mistakes and weaker reasoning. The PS5 could reach speeds of 20.9 tokens per second (tok/s) for the 27B model, with smaller models like Gemma 4 E4B hitting 61 tok/s and Llama 3.2 3B reaching 124 tok/s. The total power draw was 86W, keeping the PS5 cool at 54°C.
There are several compromises to consider. The 16GB memory limits you to running a heavily quantized 27B LLM or significantly smaller AI models, which means you won’t get the desired quality. Additionally, getting a native app of llama.cpp to run on the PS5 requires a jailbreak, which voids your warranty and could result in online game bans.
If you’re thinking about turning your PS5 into an AI server, it’s important to weigh the benefits against the risks. For most users, a dedicated mini PC with more memory and better cooling might be a more practical solution. For those interested in the technical details, the original source provides a deeper dive into the experiment and its limitations.
