A developer has found a surprising new use for Sony’s decade-old PlayStation 4: running a modern artificial intelligence model directly on the console, without a PC or an internet connection.
The open-source project, called Garlic LLM, allows a jailbroken PS4 to run Alibaba’s Qwen3.5-9B language model. Developer Leonardo Alegría says the system can generate between 31 and 39 tokens per second during regular conversations, with coding performance reaching up to 43 tokens per second.
The project is drawing attention because the original console launched in 2013, long before today’s widespread interest in locally running large language models. Its performance highlights what specialized software optimization can achieve with older gaming hardware.
How Garlic LLM Brings AI to the PS4
Garlic LLM is designed specifically for the PlayStation 4’s AMD-based Graphics Core Next (GCN) architecture. Instead of depending on conventional graphics drivers or established AI software frameworks, the application communicates with the GPU through custom low-level commands.
The developer built compute kernels using OpenCL C and compiled them with LLVM for the console’s graphics architecture. The application also creates its own PM4 command buffers, which are normally generated by a graphics driver.
This approach allows the software to send workloads directly to the GPU. The CPU handles tasks such as tokenization, the user interface, text rendering and audio, while the GPU performs the main language-model calculations.
The project does not rely on llama.cpp or a Vulkan driver port. That makes its approach different from some local AI experiments developed for newer gaming hardware.
How a 9-Billion-Parameter Model Fits Into 8GB of Memory
One of the biggest technical challenges is memory. The original PS4 has 8GB of shared GDDR5 memory, which must accommodate the operating system, graphics workloads and the AI application.
A conventional 9-billion-parameter model would require more memory if every parameter were stored at higher precision. Garlic LLM addresses this limitation through quantization, reducing the amount of memory needed to store the model.
Most of the model’s weights use 4-bit precision, while some of its largest matrices use a more aggressive 3-bit format. The developer also uses a 4-bit attention cache to keep memory requirements under control.
According to the project details, the main model occupies approximately 4.5GiB. Its context window is configured for 16,000 tokens, considerably smaller than the model’s native 256,000-token capacity.
The complete application package is approximately 5.66GB and includes the language model and additional speech-related components.
PS4 AI Performance Reaches 43 Tokens Per Second
Garlic LLM’s reported performance is particularly notable for hardware from 2013. The developer lists regular text-generation speeds of approximately 31 to 39 tokens per second, with code generation reaching up to 43 tokens per second.
A major part of the performance improvement comes from speculative decoding. Qwen3.5 includes a multi-token prediction feature that allows the system to draft up to two tokens ahead before the main model verifies them.
This reduces the amount of work required during text generation and helps the older GPU deliver faster responses. The reported speeds are developer-provided figures, however, rather than results from a standardized independent benchmark.
The project also offers a useful example of how model compression and hardware-specific optimization can make local AI possible on devices that were never designed to run modern language models.
Voice Conversations and Offline Dictation Are Also Supported
Garlic LLM goes beyond basic text conversations. It includes additional models that allow the console to process spoken input and generate spoken responses.
For voice output, the application uses Chatterbox speech models, supporting seven languages: English, Spanish, Portuguese, Russian, Italian, French and German.
Voice conversations run at a lower generation speed of approximately 22 tokens per second because speech generation adds extra processing requirements.
The project also includes NVIDIA’s Parakeet TDT 0.6B v3 model for speech recognition and dictation. Its listed language support covers 25 European languages, allowing users to speak prompts instead of typing everything with a controller.
All these features are designed to operate locally. Once the software and models are installed, an internet connection is not required to use the application.
What PS4 Owners Need to Run Garlic LLM
The application is not designed for an ordinary, unmodified PlayStation 4. It requires a console running GoldHEN, a homebrew enabler used on exploited PS4 systems.
The developer’s reported testing includes a base PS4 and PS4 Slim running firmware 11.52. The official project page also lists user reports involving additional firmware versions and PS4 Pro systems, although results can vary depending on the console and software configuration.
The installation package is available through the project’s Hugging Face repository. Users need approximately 6GB of free storage, with around 12GB potentially required if the package is first copied to the console’s internal drive.
GoldHEN installation and the use of homebrew software carry risks. Garlic LLM itself does not contain an exploit, and users should understand the implications of modifying their console before attempting to install it.
Why Garlic LLM Is Not Available for PS5
Despite the newer console offering more powerful hardware, Garlic LLM is currently designed specifically for the PS4’s GPU and system software.
The developer’s project documentation states that the application does not support PS5. Its custom compute kernels and low-level GPU communication were built around the older console’s architecture.
Separate experiments have explored running language models on PS5 hardware through different software approaches. However, those projects use their own development methods and should not be confused with Garlic LLM.
The PS4 project is therefore not simply a portable AI application that can be installed across PlayStation generations. Its performance depends on software written specifically for the hardware it targets.
An Unusual New Role for Older Gaming Hardware
The project demonstrates how older consumer electronics can take on workloads beyond their original purpose when developers have direct access to the hardware.
A PS4 was built primarily for gaming, not for running large language models. Garlic LLM shows that its GPU and memory can still support useful local AI workloads when the software is carefully optimized.
It does not turn the console into a replacement for a modern AI workstation. Its 16K context window, memory limitations and specialized installation requirements remain important constraints.
However, the ability to run a 9-billion-parameter model, process speech and generate code without relying on cloud servers makes the experiment a notable development in the growing local AI community.





