A jailbroken PS4 has been used to run Qwen3.5-9B, a 9-billion-parameter AI model, and the claim attached to it is the kind that makes modding circles stop and look twice: up to 43 tokens per second on console hardware first released in 2013.
That is why the PS4 is being searched now. The benchmark result turns an old game console into an unexpected local AI machine, and it does so with a model large enough to make the speed figure matter. Forty-three tokens per second is fast enough to move this beyond curiosity and into proof-of-concept territory, at least on paper.
The setup behind the number is Garlic LLM, described as open-source software that bypasses drivers and talks directly to the 2013-era GPU. In plain terms, that means the usual console path is not doing the work here. The model is not waiting for standard support from the PS4 stack; it is being pushed through a more direct route to the graphics hardware, which is the only way a claim like this makes technical sense.
That friction point matters because the headline rests on a jailbroken system, not ordinary PS4 support. A modified console can do things a stock machine cannot, and that distinction is the whole story here. It also explains why the result belongs in the same conversation as other PS4 software and hardware workarounds, from offline play concerns to platform limits that show up in articles like Capcom Street Fighter 5 Php Security concern forces PS4, Steam offline and Sony Playstation Network Outage hits PS4 and PS5 users on July 24.
What remains unresolved is the testing setup itself. The report provides the performance claim and the mechanism, but not the full recipe behind the benchmark conditions that produced 43 tokens per second. For now, the clearest takeaway is that an old console, a jailbroken software environment and direct GPU access were enough to run a modern 9-billion-parameter model at a speed that would have sounded impossible only a short time ago.

