PS5LM is a homebrew application that enables jailbroken PlayStation 5 consoles to run language models using llama.cpp’s Vulkan backend on the console’s RDNA 2 GPU. According to the project documentation, the application loads GGUF models onto the PS5’s 11.4 GB of shared GPU memory and makes them accessible to other devices on the network via standard APIs.
The tool supports OpenAI’s /v1/chat/completions and Anthropic’s /v1/messages endpoints, allowing users to point compatible clients—including coding tools—at the console. A web-based chat interface accessible through any browser on the same network provides an alternative way to interact with loaded models. The application also displays real-time monitoring of console temperature, power consumption, GPU and CPU load, and streaming request activity on the TV screen.
According to performance benchmarks included in the documentation, the application can run models of various sizes. A Qwen 3.5 0.8B model achieves 207 tokens per second, while a larger Qwen 3.5 9B model processes 55 tokens per second. The application includes an automatic planner that reads GGUF model files and allocates the longest context and best cache settings based on available memory.
PS5LM also supports decision models through integration with Ollaya, enabling the console to answer typed questions with calibrated probabilities. The documentation lists d1 Omni 600M as an example decision model capable of processing 3 questions in 0.3 seconds, or a 720p image and 7 questions in 2.9 seconds.
Installation requires a PS5 on firmware versions 7.00 to 13.60 that has been jailbroken with specific payloads including kstuff-lite, etaHEN, and ShadowMountPlus. Models can be downloaded from Hugging Face or manually copied to the console’s internal storage, M.2 drive, or USB drive. The project is released under GPL-3.0-or-later and is not affiliated with Sony Interactive Entertainment.
Key facts
- PS5LM uses llama.cpp’s Vulkan backend to run GGUF models on the PS5’s RDNA 2 GPU with 11.4 GB shared memory
- The application serves models to network devices via OpenAI and Anthropic-compatible APIs and a browser-based chat interface
- A Qwen 3.5 9B model achieves 55 tokens per second on a PS5 Slim according to benchmarks
- The tool supports decision models through Ollaya integration for typed question answering
- Installation requires a jailbroken PS5 on firmware 7.00 to 13.60 with specific homebrew payloads
- Models can be sourced from Hugging Face or manually copied to internal storage, M.2 drive, or USB