Developer Leonardo Alegria’s Garlic LLM runs Alibaba’s Qwen3.5-9B on a jailbroken PS4’s 2013 graphics chip, fully offline, at up to 43 tokens per second. It can even talk back.
Your old PS4 might be smarter than you think. Or at least, it can be.
A new homebrew app called Garlic LLM runs a 9-billion-parameter AI model on a jailbroken PlayStation 4, with no PC and no internet connection. Developer Leonardo Alegria released it as a free, open-source project, and Tom’s Hardware covered it on Nov. 16.
The console in question came out in 2013. The AI model running on it came out this year.
Watch the Clownfish TV podcast on YouTube. Article continues below.
What is Garlic LLM?
Garlic LLM is a chatbot app for the PS4. It runs Qwen3.5-9B, an open AI model from Alibaba, directly on the console’s own graphics chip.
Alegria describes it simply on the project page: a 9-billion-parameter language model running on a PS4, offline. It can chat by text, listen to you through a speech-to-text model, and answer out loud in seven languages. According to the project page, it can even run the Python code it writes.

The name is a nerdy joke. “Garlic” is the name of a type of graphics memory on the PS4. Its sibling memory type is called “Onion.”
How fast is the PS4 AI chatbot?
Surprisingly fast. Alegria says text chat runs at 31 to 39 tokens per second, and code gets up to 43. Voice chat runs at about 22 tokens per second.
In plain terms, at 35 tokens per second, a 300-token answer takes less than nine seconds. That’s faster than most people can read.
Tom’s Hardware notes that these are the developer’s own numbers, and the test conditions haven’t been shared. Still, the demo shows a PS4 answering questions in real time, which is wild for a console that launched alongside Killzone Shadow Fall.
The app also uses a speed trick. Qwen3.5 can guess up to two words ahead, and the full model then checks those guesses. Alegria credits that for much of the chat speed.
How does it fit on a PS4?
The PS4 has 8GB of memory, shared between its processor and graphics chip. A 9-billion-parameter model would normally need a lot more than that.
Garlic LLM shrinks the model to about 4.5GB by storing most of its numbers with 4 bits instead of 16, and some with even less. The speech models add a little more on top. The whole package is a 5.66GB download.
It also skips the usual software layers. Instead of using a graphics driver, the app builds the low-level commands for the PS4’s AMD graphics chip by hand. That’s a rare and difficult approach, and it’s a big part of why this project stands out.
The math itself is written in OpenCL C and compiled for the PS4’s GCN graphics architecture, while plain C code handles the menus, text and audio. The app was built with OpenOrbis, an open-source toolkit for PS4 homebrew.
The tradeoff is memory for conversation history. The model normally supports 256,000 tokens of context, but on the PS4 it’s limited to 16,000.
Can you run Garlic LLM on your PS4?
Only if it’s jailbroken. The app needs a PS4 running GoldHEN, a homebrew enabler for hacked consoles. Alegria tested it on base and Slim PS4 models running firmware 11.52.
It hasn’t been tested on the PS4 Pro yet, and it won’t run on a PS5. The download is free on Hugging Face, and you’ll need about 6GB of free space to install it.
Jailbreaking comes with real risks, including voiding your warranty and potential bans from online services. The app itself doesn’t contain an exploit, but anyone installing GoldHEN does so at their own risk.
Has anyone run AI on a PS5?
Yes. A separate project called PS5LM runs AI models on jailbroken PS5 consoles. A version of it ran a 27-billion-parameter model at around 20 tokens per second last month.
That’s impressive, but the PS5 is a much newer machine. Getting a 9-billion-parameter model running smoothly on 2013 hardware is the bigger surprise. It’s the latest in a long line of “can it run” projects, and this one actually runs well. Not bad for a console most people have stuffed in a closet.
Article compiled and written by Thom Pratt (editor-in-chief), Derek Gibbs (beat editor) and the Clownfish TV newsroom.
D/REZZED is part of Clownfish TV. For more news, views, and rants on gaming, tech, and pop culture, subscribe at clownfishtv.com. Watch the show on YouTube at @ClownfishTV where new episodes drop daily. Subscribe to the Clownfish TV podcast on Apple Podcasts, Spotify, iHeart, and wherever else you get your podcasts.
Hat Tips:
-
Tom’s Hardware (Nov. 16): Garlic LLM speeds, memory use, requirements and comparisons.
-
Garlic LLM on GitHub (Leonardo Alegria): Project details and features.
-
Hardware Busters (Nov. 16): Release coverage.
AI Disclosure: All articles are assigned, edited and reviewed by an experienced human editor before posting. Like most modern newsrooms, WebReef Media uses a variety of AI tools to help with accuracy and efficiency. Ways AI may be used in article creation include research, grammar, fact-checking and ideation, as well as occasional AI generated or enhanced thumbnails or graphics. No articles are generated or posted without human oversight or involvement.


