A homebrew project uses the console’s GPU to run language models locally.
Summary
– A PS5LM fork reportedly runs a small AI model at about 124 tokens per second on a jailbroken PS5
– GPU acceleration and additional usable memory allow larger models to run at lower speeds
– The project requires a modified console and is not an official PlayStation feature
A developer has demonstrated a new use for a jailbroken PlayStation 5: running AI chat models directly on the console, with one smaller model reportedly generating about 124 tokens per second.
The project builds on PS5LM, a port of the llama.cpp software used to run language models. Rather than operating as a CPU-based payload, the fork runs as a native application and uses the PS5’s graphics processor through Vulkan.
The developer reports roughly 124 tokens per second with Llama 3.2 3B and 61 with Gemma 4 E4B. A heavily compressed 27-billion-parameter Qwen model reportedly reaches around 20 tokens per second. Those figures describe different models and configurations, so the headline speed does not apply to every model the console can load.
Memory access is another major change. The developer says the native app can use approximately 11.4 GiB, compared with around 6 GiB in the earlier payload approach. That gives the software more room for model data and the memory needed to maintain a conversation.
A dashboard displays generation speed, hardware usage, temperatures and memory allocation. The app also supports a browser interface, allowing another device on the same local network to communicate with the model running on the console.
The reported output speed does not tell the whole performance story. The developer acknowledges slower prompt processing on the larger model, meaning a long document or lengthy instruction can take time to process before a response begins.
This remains a homebrew experiment for jailbroken hardware. It is not a Sony update or a feature available through the standard PS5 interface, and the reported benchmarks are developer-provided rather than independently verified results.
