In this article
Garlic LLM puts a 9B model on a 2013 console
Developer Leonardo Alegría has released Garlic LLM, an open-source app that runs Alibaba's Qwen3.5-9B model on a jailbroken PS4, according to Tom's Hardware. It works offline once installed, TweakTown reports. It is free and ships as a single 5.66GB download on Hugging Face, installed through the PS4 Package Installer.
What a PS4 needs to run Garlic LLM
You need a jailbroken PS4 running GoldHEN. The developer tested on a base or Slim PS4 on firmware 11.52. Installation needs about 6GB of free storage, or 12GB if you first copy the package to the console's internal disk. The PS5 is not supported, because the app is written for the PS4's GPU and system software. The app contains no exploit, but Tom's Hardware notes that installing GoldHEN is at your own risk.
Speeds of 31 to 43 tokens per second, self-reported
Tom's Hardware gives developer-reported speeds of 31 to 39 tokens per second in text chat and up to 43 for code. With spoken replies running alongside, chat drops to about 22 tokens per second. Tom's Hardware says the prompt and test method are not given, and TweakTown notes the figures come from developer testing and user reports, not independent benchmarks.
The project credits Qwen3.5's multi-token prediction head for the chat speed. It drafts up to two tokens ahead and the full model verifies them.
Heavy quantization fits the model into 8GB
TweakTown lists the PS4 at 8GB of shared memory. It says a 9B model at FP16 would need around 18GB. Garlic LLM uses mostly 4-bit weights, with roughly 3.5-bit weights for the two largest matrices in each layer.
That brings the model to about 4.5GiB. The attention cache is 4-bit, which allows a 16K-token context against the model's native 256K. Tom's Hardware puts the total memory budget, including the voice models, at 5.23GiB.
The app talks to the GPU directly, using OpenCL C kernels compiled with LLVM for AMD's GCN architecture and its own PM4 command buffers. Tom's Hardware says it uses neither llama.cpp nor a Vulkan port, "or so it would appear". The CPU side is C code built with the OpenOrbis PS4 toolchain.
Voice, dictation and the PS4 Pro question
Spoken replies in seven languages use the Chatterbox speech model. Holding a controller button enables dictation with Nvidia's Parakeet TDT 0.6B v3, which covers 25 European languages. TweakTown adds that the app supports local MicroPython code execution.
Tom's Hardware says the PS4 Pro is "not tested yet" and that the developer is asking for reports on GitHub. TweakTown says the developer's website lists user-reported PS4 Pro results of up to 56 tokens per second in an experimental native mode. The coverage we have does not say how the lower-bit quantization affects answer quality, or when the app was released.
The PS5LM fork is the nearest comparison
Tom's Hardware says the name comes from "Garlic", the GPU-direct memory type, and names the PS5LM fork as the closest comparison, which runs llama.cpp natively on PS5 through a RADV port called PS5_Vulkan. On that fork, Tom's Hardware lists Qwen3.8-27B at about 20 tokens per second, Gemma 4 E4B at 61 and Llama 3.2 3B at 124.
Sources
- tomshardware.com/video-games/playstation/jailbroken-ps4-runs-9-billion-pa…
- tweaktown.com/news/114709/jailbroken-ps4-runs-a-9-billion-parameter-ai…
Image: Photo by Milad Fakurian on Unsplash
This story was produced with AI assistance from the sources listed above and fact-checked against them before publication. Read our editorial methodology and the corrections log. Spotted an error? Email [email protected].



Discussion