# Garlic LLM Runs Qwen3.5-9B on a Jailbroken PlayStation 4

> Leonardo Alegría's open-source Garlic LLM runs Alibaba's Qwen3.5-9B on a jailbroken PlayStation 4 at up to 43 tokens per second.

- Published: 2026-11-17T05:38:59Z
- Score: 8/10
- Tags: ps4, exploit, ai
- Port / platforms: PS4 -> PS4 (Homebrew)
- Web page: https://emu-radar.customfw.xyz/news/garlic-llm-ps4-ai-inference/
- Sources: https://news.lavx.hu/article/jailbroken-ps4-runs-9-billion-parameter-ai-model-at-up-to-43-tokens-per-second

Developer Leonardo Alegría released Garlic LLM, an open-source homebrew app that runs Alibaba's Qwen3.5-9B large language model on a jailbroken PlayStation 4. The application bypasses standard driver layers to talk directly to the 2013-era Graphics Core Next GPU using hand-built PM4 command buffers and OpenCL C compute kernels.

Running a nine-billion-parameter model on hardware with eight gigabytes of shared memory requires aggressive quantization. Garlic LLM packs the model into roughly 4.5GiB by using 4-bit weights alongside 3-bit weights for the two largest matrices in each layer. A 4-bit KV cache allows a 16,000-token context window. Combined with Nvidia's Parakeet dictation model and the Chatterbox speech model for voice output, the full package occupies about 5.66GB.

On a base or Slim PS4 running firmware 11.52 with GoldHEN, text chat reaches 31 to 39 tokens per second, while code generation peaks at 43 tokens per second. Spoken replies run at roughly 22 tokens per second. The app leverages the model's multi-token prediction head to draft up to two tokens ahead via speculative decoding.

1. Ensure your console is running firmware 11.52 with GoldHEN installed.
2. Download the 5.66GB package file from Hugging Face.
3. Install the package using the PS4 Package Installer, ensuring you have about 6GB of free space on the internal disk, or 12GB if copying the package to the internal disk first.

Garlic LLM runs compute kernels written in OpenCL C and compiled with LLVM for AMD's Graphics Core Next architecture.
