A new open-source project called PS5LM turns a jailbroken PlayStation 5 into a private, local AI server. It ports llama.cpp, the popular C++ inference engine for large language models, to run directly on the PS5’s GPU, letting the console load and serve any compatible GGUF model to every device on your home network without a PC, a cloud account, or an internet connection.

The project matters because it is one of the more unusual demonstrations of the ongoing push to run AI models locally, on hardware never designed for the job. It also reflects how far the console jailbreaking and homebrew community has advanced in porting serious software stacks, not just games or emulators, onto locked-down consumer devices.

What PS5LM actually does

PS5LM loads a GGUF-format model, the file format used by llama.cpp, onto the PS5’s graphics processor and runs inference locally. Once running, it exposes the model over the network so that a laptop, phone, or coding tool can send it requests, much like a self-hosted version of a cloud AI API. The project ships with a TV-friendly interface featuring Home, Models, Console, API, and Settings screens, so the console itself can be used to manage downloads and configuration directly from a couch controller setup.

The suggested starting model is Qwen 3.5 9B, with the project noting it has demonstrated the setup answering requests from a laptop on a PS5 Slim. The server also exposes OpenAI and Anthropic-compatible APIs, meaning existing tools built against those APIs can point at the PS5 instead of a cloud provider with minimal changes.

How it is built, technically

Under the hood, PS5LM combines several existing open-source building blocks rather than reinventing them. It uses a homebrew UI toolkit for the PS5 interface, llama.cpp for the actual model inference, and ggml-vulkan, llama.cpp’s Vulkan backend, to talk to the console’s GPU through RADV, the open-source Vulkan driver for AMD graphics hardware. The build process happens on Linux. This combination is notable because it means the project did not need Sony’s official software development kit: the repository explicitly states it contains no Sony SDK files, keys, firmware, or model weights, keeping it within the legal grey zone that homebrew projects typically occupy.

Running PS5LM requires a console on firmware versions 7.00 through 13.60 that has already been jailbroken, along with a set of payloads loaded beforehand. Updating the software is as simple as copying a new folder over the old one, since settings, downloaded models, and configuration are stored separately in a persistent data directory, so they survive updates.

Decision models and the Ollaya connection

Beyond standard text-generating language models, PS5LM has added support for what it calls decision models, through a feature it attributes to a system named Ollaya. The pitch for Ollaya is that it runs open decision models on a PC the way Ollama, the well-known tool for running LLMs locally, runs language models. A decision model, as described in the project, reads a given state and answers typed questions in a single forward pass without generating any text, returning calibrated probabilities instead of prose, reportedly in a fraction of a second.

PS5LM brings this same concept to the console, running models such as d1 Omni 600M, d1 3B, and Clef Flash, and answering requests through a /v1/systemone API endpoint that mirrors Ollaya’s interface. The project notes that any GGUF file carrying llama.cpp’s decision metadata will automatically be recognized and labeled as a decision model by the app. This is a more specialized use case than typical chatbot-style inference, aimed at applications that need fast, structured yes-or-no or classification-style answers rather than generated text.

How this differs from typical local AI setups

Most local LLM tools, including Ollama itself, are designed to run on a desktop or laptop GPU, or increasingly on specialized AI hardware. PS5LM’s distinguishing feature is the hardware target: a game console that millions of households already own, repurposed as a dedicated, always-available inference server. Because the PS5 can run headless once the model is loaded, it does not need a monitor, keyboard, or active use as a gaming device to serve requests to other machines on the network.

This also differs from cloud-based AI services in an important way: everything runs locally, with no account creation, no API billing, and no data leaving the home network. For developers already using coding assistants, the project includes documentation for integrating with tools like opencode, Codex, and Claude Code, treating the PS5 as a drop-in local backend for AI-assisted coding workflows.

What it means for developers and hobbyists

For anyone who already owns a jailbroken PS5, PS5LM offers a novel way to extract extra utility from otherwise idle console hardware, turning it into a private model server. For developers interested in self-hosted AI, it adds one more option to a growing ecosystem of local inference tools, with the caveat that the entry requirement, a jailbroken console on a specific firmware range, is a significant barrier and carries the usual risks associated with modifying console firmware, including warranty and online service implications.

The project is licensed under GPL-3.0-or-later, while llama.cpp itself retains its MIT license and the Manrope font used in the interface keeps its SIL Open Font License. The project is explicitly unaffiliated with Sony Interactive Entertainment, and PlayStation and PS5 remain Sony’s trademarks.

Frequently asked questions

Do I need a PC to use PS5LM? No. Once the model is loaded onto the console, PS5LM is designed to run without a connected PC, serving requests directly to other devices on the network.

What PS5 firmware versions are supported? The project supports jailbroken consoles running firmware 7.00 through 13.60, with specific payloads that need to be loaded beforehand.

Can PS5LM run any AI model? It supports GGUF-format models compatible with llama.cpp, including both standard language models like Qwen and specialized decision models that carry llama.cpp’s decision metadata.

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