A Jailbroken PlayStation 5 Is Now Running a 27-Billion-Parameter AI Model

An open source project called PS5LM has turned a jailbroken Sony console into a working local AI server, according to Startup Fortune. Plug a PS5 into the same Wi-Fi network as your laptop, open a browser, and you can chat with a 27-billion-parameter language model running entirely on hardware most people bought to play games.

The project is built by a developer who goes by cobanov on GitHub and ports llama.cpp, the widely used open source inference engine behind most local AI chatbots, directly onto a jailbroken PS5. The headline number: Qwen3.8-27B, a large open-weight model from Alibaba's Qwen team, generates text at roughly 20 tokens per second on the console's GPU, and the report stresses that this is neither a typo nor a simulation on a hidden PC.

The hardware explains why it works. The PS5 ships with an AMD GPU carrying 16GB of GDDR6 memory and 448 GB/s of bandwidth, the kind of spec sheet that matters for large language models, since token generation is bound by memory bandwidth more than raw compute and that bandwidth sits idle whenever a game is not loaded.

Getting there required stacking several independent PS5 hacking projects. PS5LM uses the relapse browser exploit, which works on firmware versions 7.00 through 13.60, to run code on the console, then loads a payload built with the open source ps5-payload-dev SDK. The PS5's OS is based on FreeBSD and runs on an 8-core AMD Zen 2 CPU, so standard llama.cpp code compiles with only minor changes.

The first working version did not use the GPU at all. It ran language models as a payload on the CPU alone, streaming weights straight from disk because the console could not hold large files in memory, and managed about one token a minute, which the report calls technically functional and practically useless. The breakthrough came from routing inference through the GPU, using a RADV-based Vulkan driver ported to talk to the PS5's silicon, with a native app running llama-server over a single device heap of about 11.44 GiB.

On performance, Qwen3.8-27B quantized to about 2 bits per weight reached 20.3 tokens per second at a 32k context window, and a slightly different quantization hit 20.7 tokens per second at 64k. Smaller models move faster: Gemma 4 E4B reportedly runs at 61 tokens per second and Llama 3.2 3B at 124 tokens per second, against 14 to 16 tokens per second for a well-specified laptop running a mid-range AMD APU on Qwen3.5-35B.

PS5LM supports every architecture and quantization level llama.cpp already handles, covering Qwen, Llama, Mistral and Gemma from 1-bit weights up to full 8-bit precision. It is released under the GPL-3.0 license and states that it only runs on a PS5 you actually own, touching no game files, licensed firmware or Sony code.

What matters most is the hardware already sitting in living rooms: tens of millions of PS5s carry a GPU that, on paper, can handle a 27-billion-parameter model, so the scarce ingredient is not silicon but the unlock. Two years of open-weight releases mean running capable local AI may need only curiosity, a browser exploit and a model file from Hugging Face.