Developer runs 9-billion-parameter AI model offline on PlayStation 4 at up to 39 tokens per second

An independent developer has turned the 2013 PlayStation 4 into an offline AI chatbot. The Garlic LLM project, created by Leonardo Alegría, runs the Qwen3.5-9B model (9 billion parameters) directly on the console's AMD GCN GPU, without Linux, without an external PC and without an internet connection while in use.

How it works

The PS4 provides no compute API for homebrew and no machine-learning libraries. Alegría wrote custom OpenCL C kernels for the GCN architecture and communicates with the GPU through hand-written PM4 command buffers.

To fit in the 8 GB shared GDDR5 memory, the model uses custom quantization: most weights in 4-bit and the largest MLP matrices in 3-bit, occupying about 4.5 GiB. Context reaches 16K tokens.

With speculative decoding via Qwen3.5's own Multi-Token Prediction (MTP) head, speed rises from around 22 to 31–39 tokens per second in chat and up to 43 tokens/s for code generation. Tested on base/Slim PS4 with firmware 11.52 (reports of it working on 14.00).

Features

  • Offline text chat.
  • Voice conversations in 7 languages (including Portuguese), with TTS via Chatterbox.
  • Dictation in 25 European languages (Parakeet).
  • Python mode: the model writes and runs MicroPython on the console itself.
  • Think mode for harder questions.

Installation requires GoldHEN (homebrew). The full package is about 5.7 GB and is available on Hugging Face. The code is open-source under GPLv3 on GitHub.

The project is not affiliated with Sony. Running homebrew involves risks that are the user's own.

Sources

This content was produced with the assistance of artificial intelligence and editorially reviewed.

Por Grok