17:48Machine-translated from Azerbaijani by AIRead in Azerbaijani
PlayStation 4’s 2013 Graphics Chip Runs an AI Model Offline

Developer Leonardo Alegría has created the open-source “Garlic LLM” project, which runs an artificial intelligence model on a jailbroken “PlayStation 4” console. The 9-billion-parameter “Qwen3.5-9B” model works without an internet connection. Computations are performed on the console’s graphics processor from 2013. During text chats, the system generates between 31 and 39 tokens per second. A token is a small piece of text processed by the model.
The “PlayStation 4” console has 8 GB of memory shared by the CPU and GPU. The model’s regular 8-bit version requires at least 9 GB. The model was therefore reduced using 4-bit and 3-bit compression. As a result, its size was reduced to approximately 4.5 GiB. The data stored by the model while it is running is also compressed to 4 bits. This allows the system to take 16,000 tokens of text into account within the same conversation. The standard version of the model has a limit of 256,000 tokens. “Garlic LLM” is free and can be downloaded from the “Hugging Face” platform as a single file. The file is 5.66 GB in size. The application can be installed using the “Package Installer” tool on the “PlayStation 4” console and requires approximately 6 GB of free space. If the file is first copied to internal storage, the required free space increases to 12 GB. The project has been tested with “GoldHEN” on the standard and “Slim” models running system version 11.52. The application also supports voice chat. The “Chatterbox” model can provide spoken responses in 7 languages. In this mode, the main model processes approximately 22 tokens per second. Speech can also be converted to text by holding down a button on the controller. This uses the “Nvidia Parakeet TDT 0.6B v3” model, which recognizes 25 European languages. All the language and voice models together occupy approximately 5.62 GB. “Garlic LLM” works directly with the console’s graphics processor. This eliminates the need for additional software layers such as “Vulkan” or “llama.cpp.” The computation code was developed with “OpenCL C” and “LLVM” for “AMD”’s GCN graphics architecture. The CPU divides the text into segments and manages the interface, on-screen visuals and audio. The model predicts the next 2 tokens in advance. This allows the speed to reach 43 tokens per second when writing code. The average speed is 35 tokens per second. At this rate, a 300-token response is generated in less than 9 seconds. By comparison, the larger “Qwen3.8-27B” model generates approximately 20 tokens per second on the “PlayStation 5” console. The “PS4 Pro” has not yet been tested. The “PlayStation 5” does not work with the application either. The application does not include a tool for jailbreaking the console. The user is responsible for installing the “GoldHEN” software.
