A jailbroken PS5 can run a 27-billion-parameter AI model at 20 tokens per second
A homebrew project runs llama.cpp on the console's GPU and can serve answers to other devices on a home network. It is unofficial and needs a hacked PS5.
At a glance
- The PS5LM Plus project runs a heavily compressed Qwen 3.8 27B model at about 20 tokens per second on a jailbroken PS5.
- Smaller models are much faster: Gemma 4 E4B hit about 61 tokens per second and Llama 3.2 3B up to 124.
- It runs llama.cpp on the console GPU and can expose an OpenAI-compatible API to other devices on the network.
- It is a community project, not a Sony feature, and requires a jailbroken console.
Sony's PlayStation 5 has found an unexpected second job: running large AI language models.
A community project called PS5LM Plus — a fork of developer Mert Cobanov's PS5LM — runs the llama.cpp engine directly on the console's GPU. With a heavily compressed 6.8GB version of the 27-billion-parameter Qwen 3.8 model, it generated around 20.7 tokens per second with a 64K context window. A longer 7,775-token answer averaged about 19.7 tokens per second.
From one token a minute to 20 a second
The original PS5LM ran on the console's CPU and managed roughly one token per minute on the same model, streaming it from storage. Moving the work to the GPU changed everything. Smaller models fly: Google's Gemma 4 E4B reached about 61 tokens per second and Meta's Llama 3.2 3B up to 124.
How it works
The PS5 has 16GB of shared GDDR6 memory, and the application could use about 11.4GB. Models in the GGUF format load from internal storage, USB drives or an M.2 SSD. A dashboard shows speed, GPU and CPU load, memory, power draw and temperature. During tests the GPU ran at full utilisation while drawing about 86W and staying around 54–55°C.
The project can also expose an OpenAI-compatible API on the local network, so a laptop or phone can send prompts to the PS5 and get answers back.
The catches
This is a homebrew experiment, not a Sony feature. It requires a jailbroken console, which is unofficial and risky, and models must be squeezed hard to fit into the available memory. Still, it is a striking demonstration of how capable consumer hardware has become for running AI locally.
Questions people are asking
Can I run AI on a normal PS5?
Not with this project. PS5LM Plus requires a jailbroken PS5, which is unofficial, may break Sony's terms of service and can carry risks.
Why is a PS5 good for AI?
It has a capable GPU and 16GB of fast shared GDDR6 memory, about 11.4GB of which the app could use — enough for compressed mid-sized language models.
How we verified this story
GlobeHourly wrote this report in its own words after checking the facts against 3 independent sources: official statements, wire agencies and established news outlets. Read our standards.
