A jailbroken PlayStation 5 has become an unexpectedly capable local AI machine. A new fork of PS5LM, the community project that brings llama.cpp to Sony’s console, runs as a native application rather than a CPU-bound payload, unlocking the PS5’s GPU and substantially more memory. The result is a reported 20-plus tokens per second with a 27-billion-parameter model—and more than 100 tokens per second with smaller ones. Tom’s Hardware first reported the upgrade.
The project, called PS5LM Plus, is a fork of work by Mert Cobanov, whose original PS5LM port established that GGUF models could run locally on a jailbroken PS5. That first CPU-focused implementation was an important proof of concept, but the limits were severe: the repository says a 27B model streamed from disk managed roughly one token per minute on the console’s Zen 2 CPU. The native fork changes the equation by sending inference work to the AMD-based GPU through Vulkan.
According to the fork’s published benchmarks, Qwen3.8-27B reaches approximately 20.3 to 20.7 tokens per second depending on quantization and context settings. Smaller models are considerably faster: Gemma 4 E4B is listed at 61 tokens per second, while Llama 3.2 3B reaches 124 tokens per second. Those figures are developer-reported rather than independent laboratory measurements, and the large model’s prompt-processing speed is a separate consideration from its response-generation rate.
The technical breakthrough is access. A conventional payload receives about 6 GiB of memory and cannot use the GPU in the same way, while the native application can access the console’s broader memory pool and an approximately 11.44 GiB device heap. The fork relies on PS5_Vulkan’s RADV-based graphics stack, credited to Mihawk-99, to connect llama.cpp’s Vulkan backend with the PS5 GPU. That turns hardware normally hidden behind Sony’s software stack into a useful accelerator for local inference.
This is more than a terminal demonstration. PS5LM Plus provides a television dashboard showing generation speed, GPU and CPU use, power draw, temperatures, fan status, memory allocation, and the split between model data and KV cache. It can load GGUF files from internal storage, an M.2 drive, or USB storage, and the project advertises support for model families including Qwen, Llama, Mistral, and Gemma. The app also exposes llama.cpp’s web interface and an OpenAI-compatible API, allowing another device on the same network to chat with a model running on the console.
There is an important catch: this is homebrew for a jailbroken PS5, not a feature owners can activate on a retail system. The installation documentation lists additional components such as kstuff-lite, etaHEN, and ShadowMountPlus, while the project’s releases and setup process are aimed squarely at users already comfortable with unofficial console software. The performance is therefore best understood as a striking hardware experiment—not an imminent replacement for a gaming PC or a commercial AI workstation. Still, PS5LM demonstrates how much dormant capability can emerge when enthusiasts gain access to the console’s GPU, and it gives the PS5 a surprisingly practical second life when nobody is gaming.
