Tech

Jailbroken PS4 Runs a 9-Billion-Parameter AI Model Offline at 43 Tokens Per Second

A thirteen-year-old games console is running a modern AI chatbot with no PC and no internet connection. Garlic LLM, a free open-source homebrew app from developer Leonardo Alegría, runs Alibaba’s Qwen3.5-9B language model entirely on a jailbroken PlayStation 4, and the project shot to the top of Google’s trending searches in the United States on Tuesday, 17 November 2026. The reported speeds — 31 to 39 tokens per second in chat, and up to 43 when the model writes code — come from the developer, with no independent benchmark yet published.

Garlic LLM on PS4 — Quick Reference Summary

ProjectGarlic LLM, free open-source homebrew app
DeveloperLeonardo Alegría
ModelAlibaba Qwen3.5-9B (9 billion parameters)
ConsoleJailbroken PlayStation 4 (base or Slim, firmware 11.52 tested)
Chat speed31–39 tokens per second (developer-reported)
Code speedUp to 43 tokens per second (developer-reported)
PS4 ProUp to 56 tokens per second, user-reported, experimental native mode
Model size on diskAbout 4.5GiB after quantization
MemoryPS4’s 8GB unified GDDR5
Connection neededNone — fully offline
Download5.66GB package via Hugging Face
LicenceGPLv3 source code on GitHub

A 2013 console running a 9-billion-parameter chatbot

The PlayStation 4 launched in 2013 with an eight-core AMD Jaguar processor, a Radeon-based graphics chip with 18 compute units, and 8GB of unified GDDR5 memory shared between the CPU and GPU. It was built to play games, not to run large language models, and it offers homebrew developers no compute API, no machine-learning library, and nothing resembling CUDA or ROCm.

Garlic LLM works around that from scratch. The app builds its own compute kernels for the console’s GCN graphics architecture and dispatches work by writing PM4 command buffers by hand, talking to the GPU directly instead of going through a driver. The result is a chatbot that runs on the console alone, with the model, the interface, and the generation all happening on hardware Sony shipped more than a decade ago.

How 9 billion parameters fit into 8GB of memory

A 9-billion-parameter model stored at FP16 precision would normally need around 18GB of memory for its weights alone, more than twice what the PS4 has in total. Garlic LLM gets around that with custom mixed 3-bit and 4-bit quantization, compressing the model weights down to roughly 4.5GiB so they fit inside the console’s 8GB of shared memory.

Speed comes from a second trick. Qwen3.5 ships with a multi-token prediction head, and Garlic LLM uses it for speculative decoding, drafting up to two tokens ahead and checking them in a single pass. Earlier builds of the project ran at roughly 22 tokens per second before that optimisation was added. The 16K-token context window keeps conversations usable, and the app also supports voice features and MicroPython, although speech synthesis slows generation down.

What the speed figures actually mean — and their limits

The headline number needs careful reading. The 43 tokens per second figure is a peak reached during code generation, while ordinary chat lands in the 31 to 39 tokens per second range. Both figures are the developer’s own numbers: no source describes the prompts or the method behind them, so they should be treated as developer-reported rather than an independent benchmark.

The same caution applies to the PS4 Pro. The developer’s project page lists user-reported results of up to 56 tokens per second on the Pro using an experimental native mode, but the Pro has not been tested by the developer, who is asking users for reports on GitHub. Some coverage has suggested the speeds beat certain PCs, yet no like-for-like benchmark supports that comparison, and results will vary with hardware, settings, and task. Tokens per second is also not directly comparable across different models, quantization levels, and software.

What you need to run it, and the jailbreak catch

Garlic LLM only runs on a jailbroken PS4 with the GoldHEN homebrew enabler installed, and the app itself contains no exploit — the jailbreak is a separate step, installed at the user’s own risk. The developer tested it on base and Slim consoles running firmware 11.52, and the package is a single 5.66GB download from Hugging Face, installed through the PS4’s Package Installer. Around 6GB of free storage is needed, or 12GB with certain installation methods.

The PlayStation 5 is not supported, because the software is written specifically for the PS4’s GPU and system software. A separate project, PS5LM, runs language models on the PS5 using llama.cpp and a Vulkan-based route to that console’s graphics processor. The name Garlic LLM comes from “Garlic”, the name of the memory path the PS4’s GPU uses to access system memory — its sibling memory type is called “Onion”.

Why a PS4 chatbot matters beyond the novelty

The project landed at a moment when offline, on-device AI is drawing serious attention, and it shows how far aggressive optimisation can stretch old hardware. A console that millions of households already own can now run a capable open-weight model with no cloud account, no subscription, and no data leaving the living room.

That is also why the story travelled so fast. After VGTimes picked it up over the weekend, Tom’s Hardware covered it on Monday, 16 November, and “ps4” became the fastest-rising Google search in the United States on Tuesday. For PlayStation owners, the timing is pointed: Assassin’s Creed Valhalla’s latest update shows Sony’s older console still getting new software in 2026, and Garlic LLM shows the same box can now hold a conversation too. The source code is published on GitHub under the GPLv3 licence, so other developers can build on the approach — or test whether those speed figures hold up on their own consoles.

Frequently asked questions

Q: What is Garlic LLM? A: Garlic LLM is a free, open-source homebrew app by developer Leonardo Alegría that runs Alibaba’s Qwen3.5-9B language model offline on a jailbroken PlayStation 4.

Q: How fast does Garlic LLM run on a PS4? A: The developer reports 31 to 39 tokens per second in chat and up to 43 tokens per second when generating code. These are developer-reported figures with no independent benchmark published yet.

Q: Does Garlic LLM need the internet? A: No. The model runs entirely on the console, with no PC and no network connection required once installed.

Q: Do I need a jailbroken PS4 to use Garlic LLM? A: Yes. The console must be running the GoldHEN homebrew enabler. Garlic LLM contains no exploit, and installing GoldHEN is at the user’s own risk.

Q: Does Garlic LLM work on PS4 Pro or PS5? A: The developer tested base and Slim consoles on firmware 11.52. PS4 Pro results of up to 56 tokens per second are user-reported and untested by the developer, and the PS5 is not supported.

Q: Where can I download Garlic LLM? A: The project is distributed as a 5.66GB package on Hugging Face, with source code on GitHub under the GPLv3 licence.

Julian Vance

Julian Vance is a senior technology analyst and engineering correspondent covering the evolution of consumer electronics and industrial tech. His reporting cuts through corporate jargon to deliver data-backed hardware and software evaluations.

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *

Back to top button
×

Stay Updated!

Get our latest news and updates first on Google.

Add as a preferred source on Google