Open-source developer Leonardo Alegria has introduced the Garlic LLM project, which allows running the local Qwen3.5-9B language model on a jailbroken PlayStation 4 console. The program works fully offline on the 2013 graphics processor, generating text at a rate of 31–39 tokens per second, and when writing code it speeds up to 43 tokens per second.

Without drivers or intermediaries: how this is arranged

The main feature of Garlic LLM is direct access to the console's "hardware," bypassing standard graphics drivers or popular engines such as llama.cpp or Vulkan ports (which are used, for example, by enthusiasts on the PS5), Tom's Hardware reports.

The project was named after the Garlic memory region, to which the PS4 graphics chip has direct access (the paired system memory region is called Onion).

The compute cores are written in OpenCL C and compiled using LLVM specifically for the AMD Graphics Core Next (GCN) architecture. Instead of a graphics driver, the program independently forms low-level PM4 command buffers for the GPU, while the central processing unit (CPU) in pure C is responsible for the tokenizer, interface, text rendering, and sound.

Since the CPU and GPU share the console's 8 GB of RAM, the developer used combined compression. Most of the model's weights are compressed to 4 bits (Q4), and the two largest matrices in each layer to 3 bits (Q3). This made it possible to reduce the size of Qwen3.5-9B to 4.5 GiB (instead of 9 GB with 8-bit representation) and fit a context window of 16,000 tokens.

To speed up operation on 13-year-old hardware, speculative decoding is used: the model predicts up to two tokens ahead, then verifies them. Thanks to this, a 300-token response is generated in less than nine seconds.

Voice communication and console requirements

In addition to text chat, Garlic LLM supports full-fledged voice interaction with local speech recognition and synthesis:

  • pressing a button on the gamepad activates dictation using the Nvidia Parakeet TDT 0.6B v3 model (0.33 GiB), which supports 25 European languages;
  • the assistant can respond by voice in seven languages thanks to the Chatterbox speech model (in this mode, the generation speed is about 22 tokens per second).
  • The total package size is 5.66 GB — it is freely available on the Hugging Face platform and is installed via the standard Package Installer on consoles with the GoldHEN exploit activated. The developer tested the program on a base PS4 (Slim) with firmware 11.52. Operation on PS4 Pro has not yet been verified (the author is collecting feedback on GitHub), and the application is incompatible with PS5 due to the specific GPU and system software architecture of the fourth PlayStation.