PS5LM Runs Llama.cpp AI Models on Jailbroken PS5 Consoles

A new open-source software project called PS5LM enables jailbroken PlayStation 5 consoles to act as local artificial intelligence servers. Building upon the llama.cpp inference framework and its Vulkan backend, the tool runs GGUF-formatted models natively on the console's RDNA 2 graphics processor. The system allocates 11.4 GB of shared memory to host models and their context windows without requiring a connected PC or cloud services.

PS5LM serves models across a local network on port 8081, providing endpoints compatible with OpenAI's /v1/chat/completions and Anthropic's /v1/messages, enabling integration with coding tools such as Claude Code, opencode, and Codex. Additionally, it supports TypeSafe's /v1/systemone API via Ollaya integration for structured decision models. These specialized models return typed classifications and calibrated probabilities in a single forward pass without generating text, processing three text questions in 0.3 seconds or analyzing a 720p image with seven questions in 2.9 seconds.

The application works on PS5 consoles running firmware versions 7.00 through 13.60 that have been jailbroken. Setup requires deploying the application folder PPSA99575 to /data/homebrew over FTP, alongside payloads like ShadowMountPlus to add the app to the home screen and ps5-exporter to log power, temperature, and hardware load. Users can store models on internal storage, an M.2 drive, or an external USB drive, with direct download support from Hugging Face.

PS5LM includes a custom TV interface displaying real-time system metrics—including CPU, GPU, temperature, and power usage over two-minute and one-hour windows—and live request activity. It supports multimodal vision models like Gemma 4, MiniCPM-V, and d1, as well as decision models such as d1 Omni 600M, d1 3B, Clef Flash, OpenJev, and standard language models like Qwen 3.5 9B.