A 9-billion-parameter chatbot now fits on an old jailbroken PS4
Developer Leonardo Alegría has released Garlic LLM, an open-source project that runs the 9-billion-parameter Qwen3.5-9B model offline on a jailbroken PS4. The console must be running GoldHEN, a homebrew enabler for exploited consoles; the app itself contains no exploit. The project reports text chat at 31 to 39 tokens per second and code generation of up to 43. Voice conversation runs at about 22 tokens per second, with spoken replies in seven languages through the Chatterbox speech model. Dictation uses Nvidia's Parakeet TDT 0.6B v3, which supports 25 European languages and is triggered by holding a controller button.
The download is a single 5.66GB package on Hugging Face, installed with the PS4 Package Installer. It needs about 6GB of free space, or 12GB if copied to internal disk first. To fit the console's 8GB of shared CPU and GPU memory, the model is compressed to about 4.5GiB using 4-bit weights, with 3-bit weights for the two largest matrices per layer. The KV cache is also 4-bit, which allows a 16K-token context against the model's native 256K. The app skips graphics drivers, using OpenCL C compute kernels compiled with LLVM for AMD GCN and its own PM4 command buffers. It uses neither llama.cpp nor Vulkan. The project credits Qwen3.5's multi-token prediction head, which drafts up to two tokens ahead while the full model verifies, for much of the speed.
Why it matters: a capable language model with speech in and out, running on a console years old and with no cloud connection, shows that useful intelligence does not require new hardware or a subscription. Cheap secondhand devices become a possible route to knowledge and spoken-language access.
Caveats: the speed figures are the project's own claims, and the prompt and test method are not given. It was tested on a base or Slim PS4 on firmware 11.52. PS4 Pro is untested, and PS5 is unsupported. GoldHEN installation is at the user's own risk, and the 16K context is much shorter than the model's native window.
Source: Tom's Hardware
Many Minded