Llama.cpp on a jailbroken PS5: run any GGUF model natively on the console

What it claims

PS5LM turns a jailbroken PS5 into a LAN model server that runs with no PC, cloud or account. Any client pointed at http://<console>:8081 can use any model name, and the loaded model answers. The page's demo shows Qwen 3.5 9B serving a laptop from a PS5 Slim, but gives no throughput figure.

How it works

The stack is C++ on ps5-homebrew-ui, with llama.cpp/ggml-vulkan running on a RADV Mesa driver from the PS5_Vulkan project. The driver patches ship in patches/mesa, and the app is built on Linux.

Decision models (d1, Clef, OpenJev) read a "state" and answer typed questions in a single forward pass, without generating text. A GGUF counts as a decision model when it carries llama.cpp's decision metadata. The TV screen streams every request live, showing which device asked, the prompt and the reply as it is written. A Console page charts temperature, power, GPU load and CPU load over the last two minutes or the last hour.

How to use it

1. Download ps5lm-app-vX.Y.Z.zip from the latest release and unzip it.
2. Copy the PPSA99575 folder to /data/homebrew over FTP. ShadowMountPlus adds the app to the home screen.
3. Get a model in one of three ways: - Use Models > Suggested in the app. - Copy a .gguf to /data/PS5LM/models. - Copy a .gguf to PS5LM/models on a USB drive.

Downloads can be saved to internal storage, the M.2 drive or USB.

1. Load the model. When Home says Ready, use the browser chat or an OpenAI/Anthropic client.

To update, copy the new folder over the old one. Settings, models and downloads persist in /data/PS5LM. Setup for coding tools (opencode, Codex, Claude Code) is in docs/CLI.md.

The page gives this decision-API example verbatim:

curl http://<console>:8081/v1/systemone -H 'Content-Type: application/json' -d '{
  "state": "I was charged twice this month. Refund me or I cancel.",
  "questions": {
    "intent": {"type": "choice", "instructions": "What does the customer want?",
               "criteria": {"refund": null, "information": null, "cancel": null}},
    "angry":  {"type": "noul", "instructions": "Is the customer angry?"}
  }
}'

The suggested decision models are d1 Omni 600M, d1 3B and Clef Flash. The vision-capable models named are Gemma 4, MiniCPM-V and d1, and they accept images in both chat and decisions. The /v1/systemone API is the same one Ollaya serves on a PC.

Limits

• Requirements: a PS5 on firmware 7.00–13.60, jailbroken, with specific payloads loaded. The fetched text does not include the payload list, so check the README.
• Missing information: the page gives no benchmarks, tokens/s or memory-per-model figures. These are said to be in the "Models that fit" and "Research" docs, which were not fetched.
• Licence: GPL-3.0-or-later. llama.cpp keeps its MIT licence. The repository contains no model weights, Sony SDK files, keys or firmware. The project is not affiliated with Sony.
• Network security: the page mentions no authentication on the port 8081 server. Treat it as open to your LAN.