PS5LM runs llama.cpp GGUF models on a jailbroken PS5's GPU
PS5LM is an open-source homebrew app (GPL-3.0-or-later) for jailbroken PS5 consoles. It loads a GGUF model onto the console's GPU using llama.cpp's Vulkan backend and serves it to laptops, phones and coding tools on the local network. According to the project README, it needs no PC while it runs, no cloud and no account. A demo video caption shows Qwen 3.5 9B answering requests from a laptop on a PS5 Slim.
The README lists what the app does. It runs on the console's RDNA 2 GPU and gives the model and its context 11.4 GB of the shared memory; the README says every architecture and quantization that llama.cpp supports comes along. It exposes OpenAI's /v1/chat/completions and Anthropic's /v1/messages (the README says Claude Code works against it), plus a chat page in any browser. A client points at http://
The interface is built for the TV. Home, Models, Console, API and Settings screens are shown on the television, and every request streams across the screen as a chat: which device asked, what it asked, and the reply as it is written. Square sends a demo prompt and Triangle opens a keyboard, so the console can be tried from the couch. A Console page charts temperature, power, GPU and CPU load over the last two minutes or the last hour, beside a table of every reading.
A planner reads each GGUF file and gives it the longest context and best cache that fit the free memory. A Suggested list holds models measured on the console, and the app downloads from Hugging Face by itself, to internal storage, the M.2 drive or a USB drive. Models with an image projector (Gemma 4, MiniCPM-V, d1) accept pictures in chat and in decisions.
The newest addition is support for decision models. The README says PS5LM now runs these models on the GPU and answers the same /v1/systemone API as Ollaya, a PC tool the README compares to what Ollama is for language models. A decision model reads a state and answers typed questions in one forward pass, with no text generated, giving typed decisions with calibrated probabilities in a fraction of a second, according to the README. The example request sends a customer message, "I was charged twice this month. Refund me or I cancel.", with one choice question about intent (refund, information or cancel) and a second question about whether the customer is angry. The Suggested list includes d1 Omni 600M, d1 3B and Clef Flash, and any GGUF carrying llama.cpp's decision metadata loads the same way and is labelled a decision model.
Setup needs a PS5 on firmware 7.00 to 13.60, jailbroken, with the listed payloads loaded. The user downloads ps5lm-app-vX.Y.Z.zip from the latest release, unzips it and copies the PPSA99575 folder to /data/homebrew over FTP; ShadowMountPlus adds PS5LM to the home screen. In the app, Models > Suggested fetches a model, or a .gguf file can be copied to /data/PS5LM/models or to PS5LM/models on a USB drive. Once Home says Ready, the user opens the console's port 8081 in a browser or points an OpenAI or Anthropic client at it. Updating means copying the new folder over the old one; settings, models and downloads stay in /data/PS5LM. Use from coding tools (opencode, Codex, Claude Code) is covered in docs/CLI.md.
The credits name abhi245y, who built the native GPU app this one grew from (PS5LM Plus), llama.cpp and ggml, Mihawk-99 for PS5_Vulkan and RADV on the PS5, and several other homebrew contributors. The repository contains no Sony SDK files, keys, firmware or model weights, and the README states that PS5LM is not affiliated with Sony Interactive Entertainment.
Key facts
- PS5LM is a GPL-3.0-or-later homebrew app that runs llama.cpp on a jailbroken PS5's GPU through the Vulkan backend, with 11.4 GB of shared memory for the model and its context.
- It serves OpenAI's chat completions endpoint and Anthropic's messages endpoint on port 8081, plus a browser chat page; the README says Claude Code works against it.
- It requires a jailbroken PS5 on firmware 7.00 to 13.60; the README says it needs no PC while running, no cloud and no account.
- A new /v1/systemone endpoint serves decision models (d1 Omni 600M, d1 3B, Clef Flash on the Suggested list) that answer typed questions in one forward pass.
- The demo caption shows Qwen 3.5 9B answering requests from a laptop on a PS5 Slim; the README says it is not affiliated with Sony.
Why it matters
The project turns a games console into a private, always-available model server that speaks the same APIs as the big hosted services. The README points the idea at people who want a model at home with no PC, no cloud and no account. The addition of decision models, which return typed answers with probabilities from a single forward pass instead of generated text, makes the console useful for classification-style tasks as well as chat.
Who it affects
Mainly owners of a jailbroken PS5 on firmware 7.00 to 13.60 who want a local model server for a laptop, a phone or coding tools. Developers who already use OpenAI-style or Anthropic-style clients can point them at the console. The README also mentions opencode, Codex and Claude Code as tools covered in its CLI documentation.
How to use it
Download the ps5lm-app release zip, copy the PPSA99575 folder to /data/homebrew over FTP, and open PS5LM from the home screen (ShadowMountPlus adds it). Get a model from Models > Suggested, or copy a .gguf file to /data/PS5LM/models or to a USB drive. When Home says Ready, open http://
How solid is it
The source is the project's own README, so every capability is the maintainers' description. Code and a demo video are public. The only model shown running is Qwen 3.5 9B on a PS5 Slim, per the demo caption. No throughput figures (tokens per second), latency measurements or benchmark results are given for any model on the console. The README text also contains some garbled or unexplained names (for example 'Ollaya', 'Jev-like', 'OpenJev'), which it does not explain.
Risks and caveats
Using the app requires a jailbroken console, and the source gives no security or legal assessment of that beyond a non-affiliation notice: PS5LM is not affiliated with Sony Interactive Entertainment. The model and its context get 11.4 GB of shared memory, which limits what fits. The source does not say which quantizations or model sizes fit beyond the names on the Suggested list, nor which PS5 models other than the PS5 Slim were tested.
“It needs no PC while it runs, no cloud and no account.”
— PS5LM project README