A developer has turned a jailbroken PlayStation 5 into a small private LLM server. PS5LM, an open-source project by cobanov, runs llama.cpp on the console's GPU and serves OpenAI-style and Anthropic-style APIs to every device on your network, with no PC, cloud or account involved while it runs. It had roughly 275 GitHub stars and 65 forks when we looked.
The README's tagline sums it up: "Your PS5 runs the language model. Every device on your network uses it." For people who already own a jailbroken console, it is a surprisingly capable local-AI box. For everyone else, it is a useful data point about how much inference hardware is hiding in living rooms.
TL;DR
| Question | Answer (from the README) |
|---|---|
| What does it do? | Runs llama.cpp (Vulkan backend) on the PS5's RDNA 2 GPU and serves GGUF models over your LAN |
| How much memory for the model? | 11.4 GB of the console's shared memory for model plus context |
| Speed on a PS5 Slim | Qwen 3.5 0.8B: 207 tok/s; Qwen 3.5 9B: 55 tok/s; Qwen 3.8 27B (2-bit): about 20 tok/s |
| APIs | OpenAI /v1/chat/completions, Anthropic /v1/messages, browser chat page, and /v1/systemone for decision models |
| What do I need? | A PS5 on firmware 7.00 to 13.60, jailbroken, with several homebrew payloads |
| Licence | GPL-3.0-or-later; llama.cpp stays MIT; no weights or Sony files included |
| Is it official? | No. The project says it is not affiliated with Sony |
What PS5LM actually does
The core is llama.cpp's Vulkan backend running on the console's RDNA 2 GPU, built with a Mesa RADV driver port for the PS5. Because it is stock llama.cpp underneath, the README says every architecture and quantization llama.cpp knows comes along. You point a client at http://<console>:8081, use any model name, and the loaded model answers.
Beyond raw serving, the app adds console-specific features:
- A TV interface. Home, Models, Console, API and Settings pages, controlled with the gamepad. Every request streams across the screen as a chat showing which device asked and the reply as it is written.
- A model planner. It reads each GGUF and gives it the longest context and best cache that fit the free memory. A Suggested list holds models the author measured on the console, and the app downloads from Hugging Face to internal storage, the M.2 drive, or a USB drive.
- Console monitoring. Temperature, power, GPU and CPU load charted over two minutes or an hour.
- Vision. Models with an image projector (Gemma 4, MiniCPM-V, d1) accept images in chat and in decisions.
If you have followed llama.cpp's recent expansion beyond plain chat models, the last bullet in the README will look familiar: PS5LM also answers the typed-decision API used by decision models.
The numbers, and how to read them
Memory bandwidth and capacity decide how large a local model fits on a PS5 GPU
The README lists measurements on a PS5 Slim:
| Model | Size | Reported speed |
|---|---|---|
| Qwen 3.5 0.8B, Q4_K_M | 0.5 GB | 207 tokens a second |
| Qwen 3.5 9B, Q4_K_M | 5.3 GB | 55 tokens a second |
| Qwen 3.8 27B, UD-Q2_K_XL | 9.2 GB | about 20 tokens a second, 64k context |
| d1 Omni 600M (decision model) | 1.1 GB | 3 questions in 0.3 s; a 720p image and 7 questions in 2.9 s |
These are the developer's own numbers on one console variant, not independent benchmarks. A few things worth knowing when reading them. First, the 11.4 GB budget is the hard ceiling: it explains why the 27B model is only usable at a 2-bit quantization (9.2 GB), and why quality will trade against size. We saw the same compromise in Underdog's Saluki 27B, which squeezes a Qwen3.8 27B into 7.89 GB, where the question is always whether tool calling and reasoning survive aggressive quantization.
Second, 55 tokens per second on a 9B model is comfortable for chat and usable for small agent loops, though prompt processing speed on long contexts is not reported in the README. Third, sustained loads on a console are limited by its thermals, which is presumably why the app ships a temperature and power chart.
Decision models on a console
The README's headline "New" item is support for decision models powered by Ollaya. A decision model reads a state and answers typed questions (choice, score, yes or no) in one forward pass without generating text. The project's example sends a customer message and asks for an intent and an anger flag, returning probabilities in a fraction of a second.
If the concept is new to you, start with our explainer on what decision models are, the Ollaya runtime that popularized the API, and Liquid AI's open d1 models, which PS5LM lists in its Suggested models. The fit is natural: decision models are tiny, fast and read-only, which suits a 11 GB memory budget far better than a chat model does.
How to set it up (if you already have a jailbroken PS5)
The steps below paraphrase the README. We have not run them on hardware.
- Confirm your console is on firmware 7.00 to 13.60 and jailbroken, with these payloads: kstuff-lite, etaHEN and ShadowMountPlus (required); elfldr on port 9021 and ftpsrv on port 2121 (recommended, for telemetry and file transfer).
- Download
ps5lm-app-vX.Y.Z.zipfrom the latest GitHub release and unzip it. - Copy the
PPSA99575folder to/data/homebrewover FTP. ShadowMountPlus adds PS5LM to the home screen. - Open PS5LM, go to Models, then Suggested, and download a model, or copy a
.ggufto/data/PS5LM/modelsor a USB drive. - Load it. When Home says Ready, open
http://<console>:8081in a browser or point an OpenAI or Anthropic client at it.
A quick test from a laptop looks like any other OpenAI-compatible endpoint:
curl http://CONSOLE_IP:8081/v1/chat/completions \
-H 'Content-Type: application/json' \
-d '{"model":"any","messages":[{"role":"user","content":"Say hello"}]}'
The README says updates are done by copying the new folder over the old one, and settings, models and downloads stay in /data/PS5LM. A separate docs/CLI.md covers using the server from opencode, Codex and Claude Code.
Using it with Claude Code and other agents
Because the server speaks Anthropic's /v1/messages, the README says Claude Code can work against it. That is a notable convenience: you can try a local model in an agent loop at zero token cost. It also invites the usual caveats we raised in our piece on a free Claude Code proxy: a small quantized model on a console is not the same thing as the frontier model Claude Code is tuned for, and agentic coding with tool calls punishes weak models with loops and malformed calls. Treat it as a private, offline sandbox for light tasks, not a replacement for a hosted model.
What people are asking
Why would anyone do this? The pitch is a private, always-on model server that costs nothing extra if you already have the console, uses no PC while running, and keeps data on your network. Console GPUs are fast, and the shared memory pool is large for the price.
How does it compare with a mini PC or a Mac? Not on equal footing. Our write-ups on Apple Silicon and local inference show how much easier a supported platform is. The PS5 is a homebrew curiosity with a fixed memory ceiling, firmware limits and no vendor support.
Is the PS5 a good AI box in general? The nearby project anyps5 shows the console's hardware is a PC-class AMD design, which is why the Vulkan route works at all.
Risks and limits
- Jailbreak risk. Running unsigned code on a console can violate Sony's terms and may result in bans from online services. The jailbreak itself is outside this project and is not something we describe or link.
- Network exposure. The server listens on port 8081 with an OpenAI-style API. The README does not describe authentication in what we read, so keep it on a trusted LAN and never forward the port to the internet.
- Firmware window. Only firmware 7.00 to 13.60 is supported. Updating the console can close the door.
- Quality ceiling. At 11.4 GB for the model and its context, you are choosing between a small model at high quality and a large model squeezed to 2 bits.
- Project maturity. The repo has 80 commits, a changelog, and a roadmap, but it is a hobbyist project. Check the issues tab before relying on it.
- Licence. App code is GPL-3.0-or-later; downloaded models carry their own licences.
What this means for builders
The practical takeaway is not that you should buy a PS5 for inference. It is that llama.cpp plus Vulkan now runs on hardware nobody designed for it, which is a good sign for the portability of GGUF as a deployment format. For anyone building local-first tools, two design decisions are worth copying: a memory planner that picks the longest context that actually fits, and a pair of compatibility APIs (OpenAI and Anthropic) so existing clients just work. If you already track new local runtimes and tools, add this one to the "surprising targets" column.
Related reading
- llama.cpp decision model support
- What are decision models?
- Ollaya: Ollama for decision models
- Underdog Saluki 27B at 2 bits
- Free Claude Code proxy: terms and data risks
- anyps5: PS5 games natively on PC
- Local AI diary: Hermes agent, Qwen on a Mac Studio
- Official: PS5LM repository and llama.cpp
If you expose a local model server to agents or tools, AgentBeam is the agent security platform from the explainx.ai team that stops AI agents before they take dangerous actions.
Specs, speeds and repository details are accurate as of October 16, 2026, and come from the project README.