Garlic LLM Runs a 9-Billion-Parameter Model on a Jailbroken PS4
Developer Leonardo Alegria released Garlic LLM, an open-source app that runs a 9-billion-parameter Qwen3.5-9B model on a jailbroken PlayStation 4.
1 source 5d ago
Reporting from 1 source: GIGAZINE.
Why it matters
The PS4 has no APIs or machine learning libraries for general-purpose GPU compute from unofficial apps, so Alegria's app assembles the low-level instructions a driver would normally generate and hands them to the GPU itself, which is why a 2013 console can run a 9-billion-parameter model offline.
Key Facts
- Garlic LLM is an open-source app by developer Leonardo Alegria that runs a 9-billion-parameter model on a jailbroken PlayStation 4.
- The model is Alibaba's Qwen3.5-9B, quantized to 3 and 4 bits to fit in about 4.5 GiB.
- Garlic LLM reaches 31 to 39 tokens per second for text and up to 43 tokens per second for code.
- Running it requires GoldHEN, and Alegria confirmed it works on firmware version 11.52.
- The source code is published on GitHub under the GPLv3 license, with an install package of about 5.7 GB on Hugging Face.
It talks directly to the console's GPU, bypassing the missing drivers, and reaches 31 to 39 tokens per second for text and up to 43 for code. It needs GoldHEN and works on firmware 11.52.
The model is Alibaba's Qwen3.5-9B, quantized to 3 and 4 bits so the weights plus acceleration data fit in about 4.5 GiB. Speculative decoding, which checks several tokens ahead at once, does the rest of the speed work. The context window is 16,000 tokens.
Garlic LLM also handles voice input and conversation through the Parakeet TDT 0.6B v3 speech recognition model and the Chatterbox speech synthesis model. Voice mode costs memory, and generation drops to about 22 tokens per second. The source is on GitHub under GPLv3, with a roughly 5.7 GB install package on Hugging Face.
Written with AI by Yomimono from the 1 cited source below, including Japanese-language reporting where cited. See our editorial policy.
Add Yomimono as a preferred source on Google