PS5LM Plus runs language models on the PlayStation 5 GPU

Developer abhi245y has released PS5LM Plus, an app for running language models on the PlayStation 5’s GPU. Computation is performed locally, and users can send prompts to the model through a web interface from a smartphone or computer on the same network.

The project is based on PS5LM by Mert Cobanov, who previously ported the llama.cpp engine to the console. In the earlier version, models ran on the CPU: according to the source, Qwen generated around one token per minute. PS5LM Plus uses up to 12 GB of the console’s memory and uses the GPU. Computation runs through ggml’s Vulkan backend and a port of the RADV driver from the PS5_Vulkan project.

The tests used quantized Qwen models with 27 billion parameters. The UD-IQ2_XXS version, about 6.8 GB in size, generated 20.7 tokens per second with a 64,000-token context window. UD-Q2_K_XL occupied 9.2 GB and generated 20.3 tokens per second with a 32,000-token window. Gemma 4 E4B and Llama 3.2 3B were also tested: their speeds were 61 and 124 tokens per second, respectively.

The app supports models in GGUF format. They can be loaded from internal storage, a USB drive, an additional M.2 SSD, or Hugging Face. The screen displays generation speed, CPU load, power consumption, and temperature; an OpenAI-compatible API is available for external clients.

Running it requires a jailbroken PlayStation 5, etaHEN, kstuff, and ShadowMountPlus. The PS5LM Plus code is published under the GPL-3.0 license.

Screenshot: abhi245y / PS5LM Plus · GPL-3.0

Translation editor: Салтанов А.