Your PS5 quietly becomes the house LLM server

What it does

PS5LM is a native PS5 homebrew app that loads a GGUF model onto the console’s RDNA 2 GPU via llama.cpp’s Vulkan backend, with 11.4 GB of shared memory for model and context. It then exposes OpenAI’s /v1/chat/completions and Anthropic’s /v1/messages endpoints, so laptops, phones, and coding tools (Claude Code included) can use the console as a private model server. There’s also a TV-friendly chat interface, so you can watch requests stream in from the couch.

The interesting bit

The planner reads each GGUF and automatically assigns it the longest context and best cache that fit in free memory — the boring part is where the value is, since fitting a 27B model into a console’s shared memory budget is the actual hard problem. It also runs Ollaya-style decision models that answer typed questions (choice, score, yes/no) in a single forward pass with no text generated.

Key highlights

• Measured on a PS5 Slim: Qwen 3.5 9B at ~55 tokens/s, and Qwen 3.8 27B (UD-Q2_K_XL) at ~20 tokens/s with 64k context
• Downloads models from Hugging Face directly to internal storage, M.2, or USB
• Console page charts temperature, power, and GPU/CPU load over time
• Vision support for models with an image projector (Gemma 4, MiniCPM-V, d1)
• Any architecture and quantization llama.cpp supports works as-is

Caveats

• Requires a jailbroken PS5 on firmware 7.00–13.60 plus three required payloads (kstuff-lite, etaHEN, ShadowMountPlus)
• The 27B model needs aggressive quantization (Q2) to fit — quality tradeoffs apply
• No Sony SDK files or firmware in the repo, but you’re firmly in unsupported-hardware territory

Verdict

If you have a jailbroken PS5 and want a genuinely private, always-on model server under your TV, this is a remarkably complete answer. If your console is on current firmware, it’s a fascinating read and nothing more.