Jailbroken PS4 runs 9-billion-parameter AI model at up to 43 tokens per second
Garlic LLM runs Alibaba’s Qwen3.5-9B on a jailbroken PlayStation 4’s GPU, fully offline, at 31 to 39 tokens per second.
Developer Leonardo Alegría has created an open-source project called Garlic LLM that enables a jailbroken PS4 to run a 9-billion-parameter Qwen3.5-9B AI model offline. The model is run directly from the GPU using a custom memory type called "Garlic," which allows the model to fit within the console's limited memory. The app communicates with the GPU directly, and the model is shrunk to about 4.5GiB, allowing it to run at speeds of 31 to 39 tokens per second.
Garlic LLM is a free download that can be installed through the PS4's Package Installer, requiring around 6GB of free space, or 12GB if installed on the console's internal disk. The model is capable of voice conversation in seven languages and can also be used for dictation with Nvidia's Parakeet TDT 0.6B v3 speech-to-text model, which supports 25 European languages.
Garlic LLM uses 4-bit weights to fit the model within the console's memory, and the attention cache is also 4-bit, allowing for a 16K-token context. The developer reports that the model's performance is comparable to the PS5LM fork, which runs llama.cpp as a native app on the PS5. However, the PS5 is not supported as the app is specifically written for the PS4's GPU and system software.
Written by urgent.news from Tom's Hardware's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.