Garlic LLM pushes Ps4 to 43 tokens per second on AI

Garlic LLM ran Qwen3.5-9B on a jailbroken PS4, reaching 43 tokens per second and showing offline AI on older hardware.

Published
1 Min Read
Garlic LLM pushes Ps4 to 43 tokens per second on AI

A jailbroken PS4 ran a 9-billion-parameter AI model at up to 43 tokens per second. Garlic LLM handled the run by bypassing drivers and talking directly to the 2013-era GPU.

- Advertisement -

Garlic LLM on PS4

The model was Qwen3.5-9B. A 9-billion-parameter LLM is large enough that the result is notable on a console from 2013.

The reported speed was up to 43 tokens per second. That is fast enough to make the setup look like a working inference demo rather than a proof-of-concept that crawls token by token.

2013-era GPU performance

Garlic LLM’s direct path to the GPU is the key technical move here. It avoids the usual driver layer, which means the software is speaking more directly to the hardware than a typical consumer app would.

That setup explains how a jailbroken PS4 can handle an AI workload tied to modern model sizes. The friction point is not the model name alone, but the fact that the machine is built around a 2013-era GPU.

- Advertisement -

Offline AI on PS4

The practical takeaway is narrow but real: users interested in console hacking now have a working example of offline AI on older hardware. The next thing readers will want to know is the exact hardware modification and software setup that made 43 tokens per second possible.

Tom's Hardware framed the result as a jailbroken PS4 running Qwen3.5-9B through Garlic LLM. The unresolved detail is the full configuration behind the benchmark, because that is what would let others judge whether the same result is repeatable on similar hardware.

Advertisement
TAGGED:
Share This Article
Tech writer covering AI, cloud infrastructure, and enterprise software. Former software engineer at Google with 7 years in technology journalism.