# Garlic LLM Runs a 9-Billion-Parameter Model on a Jailbroken PS4, Offline, at Up to 43 Tokens per Second

> Source: <https://hwbusters.com/news/garlic-llm-runs-a-9-billion-parameter-model-on-a-jailbroken-ps4-offline-at-up-to-43-tokens-per-second/>
> Published: 2026-11-16 13:15:40+00:00

**A homebrew app called Garlic LLM runs Alibaba’s Qwen3.5-9B entirely on a jailbroken PlayStation 4, with no PC and no network connection. To get there, its developer skipped the graphics driver and wrote the GPU’s command packets by hand.**

Garlic LLM is a new open-source homebrew app that turns a thirteen-year-old games console into an offline chatbot. Developer Leonardo Alegría has it running Qwen3.5-9B, a 9-billion-parameter language model, on the PS4’s own GCN graphics hardware at 31 to 39 tokens per second in chat. On code it goes up to 43 tokens per second. That is faster than most people can read, on a machine whose GPU was designed in 2013 to push *Killzone* frames, not matrix multiplications.

The project was picked up by [VGTimes](https://vgtimes.com/tech-and-hardware/174706-garlic-llm-runs-qwen3.5-9b-locally-on-the-original-playstation-4.html) over the weekend and by [Tom’s Hardware](https://www.tomshardware.com/video-games/playstation/jailbroken-ps4-runs-9-billion-parameter-ai-model-at-up-to-43-tokens-per-second-open-source-garlic-llm-bypasses-drivers-to-talk-directly-to-the-2013-era-gpu) today. The most detailed account is on the developer’s own [project page](https://garlicllm.dev/).

## No compute API, so it writes its own

The hard part is not the model. The PS4 simply offers homebrew nothing to run one with: no compute API, no machine-learning library, nothing like CUDA or ROCm. The project page puts it plainly: “Garlic LLM talks to the GPU directly.” The app builds its own compute kernels for the console’s GCN architecture and dispatches them by writing PM4 command buffers manually. Those are the low-level packets a graphics driver would normally generate on the program’s behalf. The CPU side, written in C, handles the tokenizer, the interface and audio.

The name comes from the hardware too. On the PS4, “Garlic” is the memory path the GPU reads directly, with “Onion” as its cache-coherent sibling. The model’s weights live on the Garlic side of the console’s 8GB of shared GDDR5.

Fitting a 9B model into that budget takes aggressive quantization. Most weights are stored at 4-bit precision, while the two largest MLP matrices in each layer drop to 3-bit. The KV cache is 4-bit as well. The language model ends up at roughly 4.5GiB. The speed comes from speculative decoding through Qwen3.5’s built-in multi-token prediction head, which drafts up to two tokens ahead and checks them in a single pass. VGTimes reports that this lifted throughput from about 22 tokens per second to the current 31–39 range.

## More than a chat box

Alegría did not stop at text. Garlic LLM bundles Parakeet for dictation and Chatterbox for spoken replies in seven languages, so it can hold a voice conversation. There is also a Python mode that writes and runs MicroPython on the console itself, and an optional Think mode. The whole package weighs in at about 5.7GB. The [source code](https://github.com/leonardo-alegria/garlic-llm) is GPLv3, and the models keep their own licences (Apache-2.0 for Qwen3.5).

## The catches

- You need a console running GoldHEN. Garlic LLM itself contains no exploit, but the jailbreak is a prerequisite and is at your own risk.
- Testing so far covers base and Slim consoles on firmware 11.52. The PS4 Pro has not been tested, and the PS5 is not supported.
- The context window is 16K tokens. That is fine for chat and simple tool use, and short for serious programming work.
- No source describes the prompts or method behind the throughput figures, so treat them as the developer’s numbers rather than an independent benchmark.

None of this makes a PS4 a sensible AI workstation. What it does show is how much performance sits unused in old hardware once you remove the software layers that were never written for this kind of work. With memory prices where they are, that is not a purely academic point. Millions of these consoles are sitting in drawers, and every one of them has 8GB of fast memory and a GPU that, with enough stubbornness, can still do useful work.
