That PS4 gathering dust under the telly has a new trick, as long as it's jailbroken. According to Tom's Hardware, an open-source project called Garlic LLM runs Alibaba's Qwen3.5-9B language model fully offline on the console. Tom's Hardware reports 31 to 39 tokens per second in text chat and up to 43 when it's writing code.
We couldn't check the developer's own project pages directly, so every detail below comes from Tom's Hardware's write-up.
The trick: no driver at all
Developer Leonardo Alegría's app doesn't go through a graphics driver. Tom's Hardware says it talks to the PS4's GPU directly.
That's unusual. On PS5, the comparable homebrew project, the PS5LM fork, runs llama.cpp and reaches the GPU through PS5_Vulkan, a port of AMD's open-source RADV driver. Tom's Hardware notes the PS4 has no such driver layer for homebrew, so Garlic LLM appears to use neither llama.cpp nor a Vulkan port.
Instead, the heavy maths (things like matrix multiplies over 4-bit weights) is written as OpenCL C compute kernels, compiled with LLVM for AMD's Graphics Core Next architecture. The app also builds its own PM4 command buffers. Those are the low-level GPU commands a driver would normally generate for you. The CPU side, written in plain C, looks after the tokeniser, the interface, text rendering and audio. The whole thing is built with the open-source OpenOrbis PS4 toolchain.
The name is a nerdy joke, too. "Garlic" is the memory the GPU reads directly, and its sibling memory type is "Onion". The model's weights live in Garlic.
Squeezing 9 billion parameters into 8GB
The PS4 has 8GB of memory, shared between the CPU and GPU. At one byte per parameter, a 9-billion-parameter model would need at least 9GB on its own, so something has to give.
According to Tom's Hardware, most of the weights are stored at 4 bits, and the two largest matrices in each layer drop to 3-bit, about 3.5 bits once scales are counted. That mix brings the model to roughly 4.5GiB, against about 4.9GiB if everything were 4-bit. The KV cache (the model's short-term memory of the conversation) is also 4-bit, which leaves room for a 16K-token context. The model natively supports 256K.
Where the speed comes from
Qwen3.5 ships with a multi-token prediction head. Garlic LLM uses it to draft up to two tokens ahead, and the full model then checks them. Tom's Hardware says the project credits this for its chat speed on 2013 hardware. Code is the fastest case, at up to 43 tokens per second.
One caveat: Tom's Hardware points out the prompt and test method behind the 31 to 39 figure aren't given. Taking 35 tokens per second as the midpoint, it says a 300-token answer arrives in under nine seconds. For comparison, the outlet says the PS5 port ran the much smaller Llama 3.2 3B at 124 tokens per second.
It talks back, too
This isn't just a text box. Tom's Hardware reports voice conversation, with spoken replies in seven languages through the Chatterbox speech model. In voice mode the chat model drops to about 22 tokens per second. Hold a controller button and it takes dictation using Nvidia's Parakeet TDT 0.6B v3 speech-to-text model, which supports 25 European languages.
What you need
- A PS4 running GoldHEN, a homebrew enabler for exploited consoles. Garlic LLM contains no exploit of its own, and Tom's Hardware notes GoldHEN is installed at your own risk.
- Free space: the package is a single 5.66GB download from Hugging Face, installed through the PS4's Package Installer. It needs about 6GB free, or 12GB if you copy the package to the internal drive first.
- The right console: the developer tested on a base/Slim PS4 on firmware 11.52. PS4 Pro is "not tested yet", and the developer is asking for reports on GitHub. The PS5 isn't supported, because the app is written specifically for the PS4's GPU and system software.
The default temperature (how much the model's answers vary) is 0.7. Tom's Hardware notes the demo video uses 0, which makes the output predictable.
Jax's take: The chatbot is the headline, but the real story is the custom GPU path. Hand-building PM4 command buffers on a console with no homebrew driver is the kind of low-level work that could matter for other PS4 compute projects, well beyond running an LLM. As a daily assistant, a jailbroken PS4 running a quantised 9B model is a party trick. As proof the old GCN GPU can still do serious work without a driver, it's a cracker.
Reporting notes



