# PS5LM

> Runs GGUF language models on a jailbroken PS5's GPU and serves them to other devices through OpenAI- and Anthropic-compatible APIs.

PS5LM is an open-source app by Mert Cobanov that loads a GGUF model onto a jailbroken PS5's GPU using llama.cpp's Vulkan backend and serves it to laptops, phones and coding tools on the local network. It needs no PC while running, no cloud and no account. The project's latest listed release is v0.9.1.

## What It Is

PS5LM is a home model server for a jailbroken PS5 (firmware 7.00 to 13.60). The console runs the model, and other devices connect to it over HTTP at the console's address on port 8081. It exposes OpenAI's `/v1/chat/completions` and Anthropic's `/v1/messages` endpoints, plus a browser chat page. Per the project, any model name works and the loaded model answers. Using it from coding tools such as opencode, Codex and Claude Code is documented in the repository.

## The TV Interface

The app has five pages navigated with L1 and R1 on the DualSense: Home, Models, API, Console and Settings. Home streams every incoming request and its reply as a chat, showing which device asked. The Console page charts temperature, power, GPU and CPU load over the last two minutes or hour. The Models page lists models measured on the console, searches Hugging Face and downloads to internal storage, an M.2 drive or USB. A planner reads each GGUF and assigns the longest context and best cache that fit free memory. Models with an image projector, such as Gemma 4, MiniCPM-V and d1, accept images.

## Decision Models

Powered by Ollaya, PS5LM can load decision models that answer typed questions (choice, score, yes/no) about text and images in one forward pass through `/v1/systemone`, with no text generated. Suggested models include d1 Omni 600M, d1 3B and Clef Flash, and any GGUF carrying llama.cpp's decision metadata loads the same way.

## Setup Path

Installation requires the kstuff-lite, etaHEN and ShadowMountPlus payloads; elfldr and ftpsrv are recommended. Users download the release zip, copy the PPSA99575 folder to `/data/homebrew` over FTP, then get or copy a model and load it. Settings, models and downloads persist in `/data/PS5LM` across updates. The project states it is not affiliated with Sony and includes no Sony SDK files, keys, firmware or model weights.

## Measured Performance

On a PS5 Slim the project reports about 207 tokens per second for Qwen 3.5 0.8B Q4_K_M, 55 for Qwen 3.5 9B Q4_K_M, and about 20 for a 27B 2-bit model with a 64k context. It states that about 11 GB of model and context fits.

## Features
- Runs GGUF models on the PS5 GPU via llama.cpp Vulkan backend
- OpenAI-compatible /v1/chat/completions endpoint
- Anthropic-compatible /v1/messages endpoint
- Browser chat page served from the console
- Decision model support via /v1/systemone
- Live request streaming shown as chat on the TV
- Console page with temperature, power, GPU and CPU charts
- Context and cache planner that fits models to free memory
- Hugging Face search and downloads to internal, M.2 or USB storage
- Image input for models with an image projector
- DualSense-controlled five-page interface

## Integrations
llama.cpp, Ollaya, Hugging Face, Claude Code, OpenAI API clients, Anthropic API clients, etaHEN, ShadowMountPlus, kstuff-lite, PS5_Vulkan, opencode, Codex

## Platforms
MACOS, LINUX, API, CLI

## Pricing
Open Source

## Version
v0.9.1

## Links
- Website: https://ps5lm.cobanov.dev
- Documentation: https://github.com/cobanov/PS5LM/blob/main/docs/CLI.md
- Repository: https://github.com/cobanov/PS5LM
- EveryDev.ai: https://www.everydev.ai/tools/ps5lm