# Garlic LLM enables 9B parameter AI models on jailbroken PS4 via direct GPU access

> **Key takeaway:** An open-source project uses OpenCL C kernels and direct PM4 command buffers to run quantized LLMs on legacy PlayStation 4 hardware.

**Published:** 2026-11-16T13:00:00+00:00 
**Source:** Tom&#x27;s Hardware 
**Category:** open-hardware 
**Canonical URL:** https://fosswire.org/news/garlic-llm-enables-9b-parameter-ai-models-on-jailbroken-ps4-.html 

## Summary
Developer Leonardo Alegría has released Garlic LLM, an open-source project that allows jailbroken PlayStation 4 consoles to run the 9-billion-parameter Qwen3.5-9B model. By bypassing standard driver layers and communicating directly with the console's 2013-era AMD Graphics Core Next (GCN) GPU, the application achieves significant inference speeds on limited hardware.

The project utilizes the OpenOrbis PS4 toolchain and requires a console running GoldHEN homebrew enabler. To fit within the PS4's shared 8GB memory pool, the model employs aggressive quantization, including 4-bit weights and 3-bit weights for the largest matrices, alongside a 4-bit KV cache. This allows the entire pipeline, including speech-to-text and text-to-speech components, to operate within the console's memory constraints.

## Technical implementation of Garlic LLM
- **Direct Hardware Interface:** The application bypasses Vulkan and llama.cpp, instead building its own PM4 command buffers to send low-level instructions directly to the GCN GPU.
- **Compute Kernels:** Matrix multiplications and other model operations are implemented as compute kernels written in OpenCL C and compiled via LLVM for the AMD architecture.
- **Memory Optimization:** The system uses a combination of 4-bit (Q4) and 3-bit (Q3) quantization to reduce the 9B model to approximately 4.5GiB, leaving room for the Parakeet TDT 0.6B speech-to-text and Chatterbox voice models.
- **Speculative Decoding:** Performance is boosted by utilizing the Qwen3.5 multi-token prediction (MTP) head, which allows the system to draft up to two tokens ahead for verification, reaching speeds up to 43 tokens per second.
- **Software Stack:** The project is built using the OpenOrbis PS4 toolchain, with the CPU handling tokenization, text rendering, and audio in plain C.

## Why it matters
This project demonstrates the potential for running sophisticated AI workloads on legacy, specialized hardware through low-level software optimization. It highlights how open-source toolchains and direct hardware access can extend the utility of aging consumer electronics for local, offline AI applications.