Developer Leonardo Alegría released Garlic LLM, an open-source app that runs Alibaba's Qwen3.5-9B model on a jailbroken PlayStation 4. The app runs the model entirely on the console, with no PC and no network connection, using the console's 2013-era GPU. Its language generation, speech recognition and voice synthesis all run locally, and the download includes the needed models for use without an internet connection after installation.

How it works on old hardware

The PS4 offers homebrew no compute API, no machine-learning library, and nothing like CUDA or ROCm for running a model. Garlic LLM bypasses the graphics driver by building its own compute kernels for the PS4's GCN architecture and dispatching them through manually written PM4 command buffers. The model's weights sit on the Garlic side of the console's 8GB of shared GDDR5 memory, which is the memory path the GPU reads directly. The project is named after that memory, and its sibling memory type is called Onion.

The quantization work shrinks the model's primary weights to about 4.5GiB so they fit in the PS4's memory. Most weights are quantized to 4-bit, the two largest MLP matrices in each layer to 3-bit, and the KV cache to 4-bit. The 4-bit attention cache lets the app keep a 16,000-token context window within the console's memory limits. Speed comes from speculative decoding that uses Qwen3.5's built-in multi-token prediction head to draft up to two tokens ahead and check them in a single pass.

Reported performance

The developer reports text-generation speeds of 31 to 39 tokens per second on the PS4, reaching up to 43 tokens per second when generating code. In voice conversation, speed falls to about 22 tokens per second, because the additional speech models take memory otherwise available for speculative decoding, according to the developer. These figures are developer measurements and have not been verified through standardized benchmarking.

Features

Garlic LLM supports voice input and conversation with Nvidia's Parakeet TDT 0.6B v3 for dictation and the Chatterbox model for spoken replies. Dictation, enabled by holding a controller button, supports 25 European languages. Spoken replies are available in seven languages: English, Spanish, Portuguese, Russian, Italian, French and German.

A separate Python mode has the model write MicroPython programs that run directly on the console. An optional thinking mode lets the AI reason before answering, and the developer notes a reasoning response can take about half a minute.

Requirements and availability

Running the app requires a PS4 running GoldHEN, a homebrew enabler for exploited consoles. Garlic LLM itself contains no exploit, and the jailbreak is a prerequisite that users take on at their own risk. Installation needs about 6GB of free space, or 12GB if the package is first copied to the console's internal disk.

The free package is a single download of about 5.7GB on Hugging Face and installs through the PS4 Package Installer. The source code is on GitHub under the GPLv3 license, while the bundled AI models keep their own licenses. The app was built with the open-source OpenOrbis PS4 toolchain, and its repository includes installation instructions, technical documentation and model-conversion utilities.

Tested consoles

The developer mainly tested the app on a base or Slim PS4 running firmware 11.52 with GoldHEN. The PS4 Pro has not been tested, and the PS5 is not supported. The app does not support the PS5 because its GPU execution code and system integration were built specifically for the PS4. A user report in the project documentation describes it running on a PS4 Pro with firmware 12.02 at roughly 25 to 45 tokens per second.