A hacked PlayStation 5 is being used to execute a 27-billion-parameter AI model directly on the console's GPU, delivering generation speeds of about 20 tokens per second.
The work stems from PS5LM Plus, a fork of Mert Cobanov's PS5LM effort. The original project introduced llama.cpp to a jailbroken PS5 via a CPU-based payload, while PS5LM Plus adds a native app that can harness the GPU.
The experiment drew attention after @Pirat_Nation posted on X, noting that a jailbroken PS5 was running Qwen 3.8 27B locally at roughly 20 tokens per second.
Qwen 3.8 27B Runs at Around 20 Tokens Per Second
PS5LM Plus has successfully executed Qwen 3.8 27B, an AI model with about 27 billion parameters.
The console's 16 GB of shared GDDR6 memory forces heavy compression for the model to fit.
Testing employed a heavily quantized version, with the app accessing roughly 11.44 GB of shared GPU and CPU memory.
A jailbroken PS5 is now being used to run a powerful AI model directly on the console.
A project called PS5LM has managed to run Qwen 3.8 27B, an AI model with around 27 billion parameters, using the PS5’s GPU.
The PS5 has only 16GB of memory, so the AI model needs heavy… pic.twitter.com/hZCu8c7zLs
— Pirat_Nation 🔴 (@Pirat_Nation) October 14, 2026
A 6.8 GB IQ2_XXS variant achieved about 20.7 tokens per second with a 64K context window.
A 9.2 GB quantized version hit around 20.3 tokens per second, and a longer 7,775-token reply averaged roughly 19.7 tokens per second.
The project dashboard reported generation near 20.9 tokens per second.
GPU Hits 100% While CPU Usage Stays Low
During the trials, the PS5's GPU ran at full 100 % utilization, while CPU load stayed very low.
Power draw was about 86 W and temperatures hovered around 54 °C to 55 °C.
This indicates the GPU handled most of the AI workload, not the CPU.
Smaller Models Run Much Faster
Less-demanding models performed considerably quicker on the same hardware. Gemma 4 E4B reached roughly 61 tokens per second, and Llama 3.2 3B peaked at 124 tokens per second.
The gap between CPU and GPU execution is stark; Qwen 3.8 27B managed only about one token per minute when streamed from storage and run via the earlier CPU-based PS5LM setup.
PS5 Can Act as a Local AI Server
PS5LM Plus runs llama.cpp directly on the PS5 GPU and provides a dashboard showing generation speed, GPU and CPU usage, memory use, power draw, and temperatures.
It can load GGUF models from internal storage, USB drives, or an installed M.2 disk.
The tool can also expose an OpenAI-compatible API over the local network, letting another computer or phone query the model while inference runs on the PS5.
PS5LM Plus is a community homebrew effort, not an official Sony feature, and requires a jailbroken PlayStation 5.
