PS5LM is a homebrew application that loads GGUF language models onto a jailbroken PS5's RDNA 2 GPU and serves them over your local network via OpenAI and Anthropic-compatible APIs. It requires no PC, no cloud account, and no subscription. You point your laptop, phone, or coding tools at the console's IP and it answers. The performance numbers are surprisingly real. Qwen 3.5 0.8B at Q4KM runs at 207 tokens per second. The 9B variant manages 55 tok/s. Even the 27B Qwen 3.8 at aggressive Q2KXL quantization fits in the console's 11.4 GB of usable shared memory, delivering around 20 tokens per second with a 64k context window. These are not theoretical — they're measured on a PS5 Slim. The technical stack is llama.cpp running on ggml-vulkan through PS5Vulkan's RADV driver port, built on top of the ps5-homebrew-ui framework. It supports every architecture and quantization that llama.cpp knows. The project also integrates Ollaya-powered decision models — small classifiers that answer typed questions about text and images in a single forward pass without generating text. The d1 Omni 600M model handles 3 questions in 0.3 seconds or a 720p image with 7 questions in 2.9 seconds. The interface is purpose-built for a TV. Every request streams across the screen as a live chat showing which device asked, what it asked, and the reply as it's written. Square sends a demo request, Triangle opens a keyboard. A Console page charts GPU temperature, power draw, and load over time. The model planner reads each GGUF and automatically selects the longest context and best cache that fit available memory. Models download directly from Hugging Face to internal storage, the M.2 drive, or USB. The catch is the jailbreak requirement. You need firmware 7.00 to 13.60 with kstuff-lite, etaHEN, and ShadowMountPlus payloads loaded. This limits the audience to people already running homebrew on their consoles — it's not a mass-market product. Installation involves FTPing a folder to /data/homebrew. Updating means copying the new folder over the old one. The project grew from abhi245y's PS5LM Plus, which first demonstrated native GPU inference on the console. This version adds the full API server, decision model support, the TV interface, the model planner, and multi-device serving. It's GPL-3.0 with llama.cpp retaining its MIT license. No Sony SDK files, keys, firmware, or model weights are included in the repository. What makes this interesting isn't the raw capability — a $200 mini PC with a discrete GPU would outperform it. It's the proof of concept: a game console's GPU, designed for shader workloads, running general-purpose inference well enough to serve as a household AI endpoint. The PS5 sits under your TV already. It has a GPU, cooling, storage, and a network connection. PS5LM asks why that hardware should sit idle when you're not gaming.
