Garlic LLM runs a 9-billion-parameter model on a jailbroken PlayStation 4 without a network
XORit Editorial · Sources, AI assistance and corrections

The open-source homebrew app Garlic LLM runs Alibaba's Qwen3.5-9B language model on a jailbroken PlayStation 4, with no computer and no network connection. In chat the model produces 31–39 tokens per second, and on code it reaches 43. Автор проекта described it in detail on his own page, which is where the data comes from.
The main difficulty is not the model but the console: it gives homebrew neither a compute API nor machine-learning libraries such as CUDA or ROCm. The app builds the compute kernels for the GCN architecture itself and sends them to the GPU, manually assembling PM4 command buffers — the low-level packets normally generated by the graphics driver. The processor-side part is written in C and handles the tokenizer, the interface and audio.
The name comes from the console itself: on the PS4, Garlic is the name of a memory region that the GPU reads directly, and Onion is its coherent neighbor. The model weights are placed in Garlic within the shared 8 GB of GDDR5 memory.
To fit a 9-billion-parameter model into that budget, it was quantized noticeably: most weights are stored in 4 bits, the two largest MLP matrices in each layer are dropped to 3 bits, and the KV cache is 4-bit as well. The language model takes up about 4.5 GiB. The speed comes from speculative decoding via a multi-token prediction head built into Qwen3.5: it runs two tokens ahead and verifies them in a single pass. According to VGTimes, this raised throughput from roughly 22 tokens per second to 31–39.
Text is not the limit of its capabilities. The package includes dictation with Parakeet and spoken answers with Chatterbox in seven languages, a Python mode that runs MicroPython on the console itself, and an optional thinking mode. The whole package weighs about 5.7 GB, the source code is distributed under GPLv3, and the models have their own licenses: Qwen3.5 is Apache-2.0.
There are limitations too. A console with GoldHEN is required: the exploit is not in the app itself, but the jailbreak is mandatory, and it is done at your own risk. The base and Slim versions were tested on firmware 11.52, the PS4 Pro was not tested, and the PS5 is not supported. A 16k-token context is sufficient for chat and simple tools but small for serious work with code. The source does not disclose the prompts or the measurement methodology, so the speed figures should be treated as the developer's data rather than an independent benchmark.
The authors speak cautiously about practical value: there is no point in turning a PS4 into a workstation for AI. Another conclusion is telling: a lot of unused performance remains in old hardware if you remove the software layers that were never written for such tasks.
Translation editor: Салтанов А.
Source: hwbusters.com ↗