Jailbroken PS4 Runs 9-Billion-Parameter AI Model Locally via Direct GPU Bypass
Garlic LLM lets a jailbroken PlayStation 4 run large language and speech models offline on its 2013-era GPU.
Developer Leonardo Alegría released Garlic LLM, an open-source project that lets a jailbroken PlayStation 4 (PS4) run the 9-billion-parameter Qwen3.5-9B AI model offline on the console’s 2013-era GPU. The app talks directly to the hardware and bypasses standard drivers, reaching 31 to 39 tokens per second in text chat.
The project runs on a PS4 utilizing GoldHEN, a homebrew enabler for exploited consoles. Garlic LLM is a single 5.66GB download available on Hugging Face, requiring about 6GB of free space to install via the PS4 Package Installer, or 12GB if copied to the internal disk first. Testing was done on a base or Slim PS4 on firmware 11.52.
The software also supports voice conversations through Alibaba’s small open-weight Qwen3.5-9B model paired with the Chatterbox speech model for spoken replies in seven languages at roughly 22 tokens per second. Holding a controller button enables dictation using Nvidia’s Parakeet TDT 0.6B v3 speech-to-text model, which supports 25 European languages.
Unlike the PS5LM fork on PlayStation 5, which uses llama.cpp and PS5_Vulkan, Garlic LLM uses neither llama.cpp nor a Vulkan port. Instead, matrix multiplies over 4-bit weights run as compute kernels written in OpenCL C and compiled with LLVM for AMD’s Graphics Core Next architecture. The application builds its own PM4 command buffers, while the CPU handles tokenization, user interface rendering, and audio. The project is built using the open-source OpenOrbis PS4 toolchain.
How to try it
- Prepare a jailbroken PS4 running GoldHEN on a compatible firmware version like 11.52.
- Download the Garlic LLM package file from Hugging Face.
- Ensure you have at least 6GB of free space (or 12GB if transferring to the internal drive).
- Install the application using the PS4 Package Installer.


