PS5LM Project Runs Llama.cpp and Decision Models on Jailbroken PlayStation 5 Consoles
A new open-source tool enables jailbroken PS5 users to host local AI models on the console's GPU, serving them to network devices without cloud dependency.
Key points
- PS5LM utilizes the console's RDNA 2 GPU via llama.cpp’s Vulkan backend to run GGUF models.
- The application supports OpenAI and Anthropic API endpoints for local network access.
- Decision models like d1 and Clef can now process typed decisions with calibrated probabilities.
A new open-source project named PS5LM allows users with jailbroken PlayStation 5 consoles to run large language models natively on the hardware. According to a report on Hacker News, the application leverages the console’s graphics processing unit to serve AI models to other devices on the local network without requiring a separate PC or cloud connection.
The software loads GGUF format models onto the PS5's RDNA 2 GPU, utilizing up to 11.4 GB of shared memory for model weights and context. It employs llama.cpp’s Vulkan backend to support various architectures and quantization levels. This setup enables the console to function as a private model server, accessible via a browser or API clients on laptops, phones, and coding tools on the same network.
Technical Capabilities and API Support
PS5LM provides API endpoints compatible with OpenAI’s /v1/chat/completions and Anthropic’s /v1/messages, allowing tools like Claude Code to interact with the local model. Users can point clients at the console’s IP address on port 8081. The application also supports decision models, including d1, Clef, and OpenJev, which answer TypeSafe’s /v1/systemone API with typed decisions and calibrated probabilities.
The interface displays active requests on the TV screen, showing which device initiated a query and the resulting text generation. Physical controller inputs are supported, with the Square button sending demos and the Triangle button opening a keyboard for direct input from the couch. A dedicated Console page charts temperature, power, and CPU/GPU load over two-minute or one-hour intervals.
Model Management and Performance
The application includes a planner that reads GGUF files to determine the longest context and best cache settings that fit the available memory. It maintains a suggested list of models measured for console compatibility and can automatically download files from Hugging Face to internal storage, M.2 drives, or USB drives. Image-capable models such as Gemma 4, MiniCPM-V, and d1 can process images within chat and decision workflows.
Performance benchmarks cited in the project documentation show three questions answered in 0.3 seconds and a 720p image with seven questions processed in 2.9 seconds. The system operates without generating text for decision models, instead reading states and answering typed questions in a single forward pass. This mirrors the functionality of Ollaya on personal computers.
What should you do?
Security professionals should monitor network traffic for unauthorized API endpoints on consumer devices, as jailbroken consoles can expose internal services. Users interested in local AI should verify the stability of their console’s firmware before attempting modifications, as jailbreaking may void warranties or brick hardware. Organizations should ensure that local network segmentation prevents consumer devices from accessing sensitive internal resources or acting as pivot points.
Background: Large language models
Large language models are a specific type of neural network trained on text. Generative AI is the broader umbrella term for any system that creates new content, including images and code. Not all generative AI uses large language models, but most modern text generators do.
What to do and how to stay safe: PS5LM
- Verify that consumer gaming consoles are isolated on a separate VLAN or guest network to prevent them from accessing critical internal servers or data.
- Monitor local network logs for unexpected traffic on port 8081 or other common AI service ports originating from gaming hardware.
- Ensure that any jailbroken or modified devices are not connected to networks containing sensitive intellectual property or personal identifiable information.
- Review firewall rules to block outbound connections from gaming consoles to external API endpoints if local model hosting is not authorized.
Step-by-step guide: Mobile Device Security: How to Protect Data Beyond Passwords
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Does PS5LM require a jailbroken PlayStation 5 to function?
Yes, the project is designed specifically for users with jailbroken PS5 consoles who want to run private model servers at home without a PC.
Which AI models and APIs are supported by PS5LM?
It supports GGUF models via llama.cpp, OpenAI’s chat completions API, Anthropic’s messages API, and TypeSafe’s systemone API for decision models.
Can PS5LM process images alongside text?
Yes, models with an image projector such as Gemma 4, MiniCPM-V, and d1 can take pictures in chat and decision workflows.
