Developer Leonardo Alegría released Garlic LLM, an open-source homebrew app that runs Alibaba's Qwen3.5-9B model completely offline on a jailbroken PlayStation 4. The 2013 console reaches up to 43 tokens per second when generating code.
What happened?
Garlic LLM bypasses the PS4's standard graphics drivers and talks directly to the 2013 AMD GCN GPU via custom OpenCL C kernels and PM4 command buffers. To fit in the 8 GB unified memory, the model was quantized to 4 bits (with large matrices at 3 bits), ending up at about 4.5 GiB. Text generation hits 31-39 tokens/s and code up to 43 tokens/s thanks to speculative decoding. It also includes voice with recognition in 25 languages and synthesis in 7.
Why does this matter?
It shows that old hardware can still run modern LLMs with aggressive optimizations, without needing a PC or internet while running. The project is available on Hugging Face as a roughly 5.7 GB package and requires GoldHEN. Tested on base/Slim PS4 with firmware 11.52.
What changes in practice?
Users with a jailbroken PS4 can have a local offline chatbot on the console. The name "Garlic" comes from the memory area the GPU accesses directly on the PS4. It's an impressive example of low-level engineering for AI on limited hardware.
By Grok