What we know

A jailbroken PlayStation 4 has been shown running Alibaba’s Qwen3.5-9B, a 9-billion-parameter AI language model, using the open-source Garlic LLM project. The model operates fully offline and achieves speeds between 31 and 43 tokens per second. Garlic LLM bypasses the PS4’s standard drivers to communicate directly with its GPU, enabling this performance on the console’s hardware. The jailbroken PS4 runs the Qwen3.5-9B model with Garlic LLM. Token generation speed reaches up to 43 tokens per second. Garlic LLM accesses the PS4 GPU directly, skipping standard driver layers.

Why it matters

This demonstration highlights how large AI models can be adapted to run on older consumer hardware through software innovation and hardware access techniques. Running a 9-billion-parameter model offline on a PS4 shows the potential for more accessible AI applications without relying on cloud services. The Garlic LLM project’s approach of bypassing drivers to tap directly into the GPU is a novel method that could inspire similar efforts on other devices.

What is still unknown

No independent evaluations of the model’s accuracy or performance quality on the PS4 have been published. Details about which languages the Qwen3.5-9B model supports in this setup remain unspecified. There is no information on whether Garlic LLM or this PS4 implementation will be made broadly available or receive official updates.