Jailbroken PS4 runs 9-billion-parameter AI model at up to 43 tokens per second

Jailbroken PS4 runs 9- billion- parameter AI model at up to 43 tokens per second. Open- source Garlic LLM bypasses drivers to run Qwen3.5- 9B on legacy hardware.

PlayStation 4 console
PlayStation 4 console

Using a decade-old gaming console to run a large language model might seem like a cool trick, but shows how adaptable modern AI is to legacy systems without all the typical software stacks in place. An open source project shows off how we can do serious computations on what is basically computer equipment most of us have thrown out of active service.

Open-source Garlic LLM bypasses drivers to run Qwen3.5-9B on legacy hardware

A developer named Leonardo Alegría made an open source project called Garlic LLM to run the 9B parameter Qwen3.5-9B AI model on a jailbroken PlayStation 4. This is done by bypassing the standard OS drivers of the console to get it to talk directly to the 2013 Graphics Core Next (GCN) GPU. This direct interaction with the hardware means they can do things like generate text and code without needing to deal with all the other gaming or system software that would normally be involved.

PlayStation 4 console
PlayStation 4 console

To get the AI model into the tight confines of the PS4 (only 8GB of shared memory), they make some extreme 4-bit and 3-bit quantizations of the model. The final packaged AI model comes out at only about 4.5GB. The full download size is 5.66GB. The system can get 31-39 tokens per second for general text and up to 43 tokens per second when generating code.

There's also voice conversational features with the Chatterbox model as well as dictation options using the Parakeet TDT 0.6B v3 model. To use it you'll need to install the GoldHEN homebrew enabler and have at least 6GB of free storage space on your system. It's an open source project available for download worldwide on Hugging Face.

They haven't confirmed performance on the PS4 Pro as the dev hasn't tested that specific model yet. There's no info on what prompts they were using or how they tested to get these speed numbers in the source material. Just an interesting proof of concept for using AI for inference on an older and more limited gaming architecture.

Comments (0)

Add a comment

Keep it civil, relevant, and useful to other readers.