# Garlic LLM brings offline AI generation to jailbroken PS4 consoles

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

- Published: 2026-11-17T05:31:33Z
- Score: 8/10
- Tags: ps4, exploit, ai
- Port / platforms: PS4 -> PS4 (Homebrew)
- Web page: https://emu-radar.customfw.xyz/news/jailbroken-ps4-runs-garlic-llm/
- Sources: https://www.tweaktown.com/news/114709/jailbroken-ps4-runs-a-9-billion-parameter-ai-model-at-up-to-43-tokens-per-second/index.html

Running modern large language models usually requires desktop silicon, but the PlayStation 4 scene continues finding unexpected life in aging AMD hardware. Developer Leonardo Alegría released [Garlic LLM](https://github.com/leonardo-alegria/garlic-llm), a free, open-source application that runs Alibaba's Qwen3.5-9B language model entirely offline on a jailbroken console.

According to developer tests, a standard PlayStation 4 or PS4 Slim hits between 31 and 39 tokens per second during regular chats, pushing up to 43 tokens per second when generating code. An experimental native mode on the PS4 Pro reaches up to 56 tokens per second. Storing a 9-billion-parameter model locally on a machine with 8GB of unified GDDR5 memory required heavy engineering. The app uses mixed 3-bit and 4-bit quantization to shrink model weights down to approximately 4.5GB, leaving enough headroom for the operating system and inference buffers.

To squeeze performance out of the console's Jaguar CPU and Radeon-based GCN architecture, the project bypasses conventional compute APIs. It uses custom AMD Graphics Core Next compute kernels and low-level PM4 command buffers to push workloads straight to graphics hardware, alongside speculative decoding via Qwen3.5's multi-token prediction.

The application also handles voice interaction, local MicroPython code execution, and speech recognition. Running text-to-speech concurrently drops generation speeds to about 22 tokens per second.

### How to try it

1. Ensure your console is jailbroken and running GoldHEN.
2. Have approximately 6GB of free storage available for installation, or up to 12GB depending on the chosen method.
3. Download and install [Garlic LLM](https://github.com/leonardo-alegria/garlic-llm) to run the model completely offline.
