PS5LM: Running Native GGUF Language Models on Jailbroken PlayStation 5

Overview

PS5LM is a homebrew project that transforms a jailbroken PlayStation 5 into a dedicated, private language model server. By leveraging the console's RDNA 2 GPU via the Vulkan backend of llama.cpp, the application allows users to run GGUF-formatted models natively, serving them to other devices on the network through standard OpenAI and Anthropic-compatible APIs.

Technical Capabilities

Designed for users with jailbroken consoles on firmware versions 7.00 through 13.60, PS5LM utilizes approximately 11.4 GB of shared memory for model weights and context. The application provides a full interface directly on the TV, including real-time monitoring of console thermals, power consumption, and GPU/CPU load.

Key features include:

  • API Compatibility: Exposes /v1/chat/completions and /v1/messages endpoints, allowing integration with tools like Claude Code or standard browser-based chat interfaces.
  • Decision Model Support: Integrates with the Ollaya ecosystem to run decision models, enabling typed, probability-calibrated responses for state-based queries without generating full text.
  • Model Management: Includes a built-in downloader for Hugging Face models and supports local storage via internal drives or USB.
  • Performance: On a PS5 Slim, the system achieves approximately 55 tokens per second for a Qwen 3.5 9B (Q4_K_M) model, and roughly 20 tokens per second for a 27B model (UD-Q2_K_XL).

Community Perspectives and Use Cases

Community discussion surrounding the release highlights both the technical achievement and the practical limitations of the hardware. While some users view the project as a "cool achievement" with limited utility due to VRAM constraints, others see potential for long-running, low-intelligence tasks.

Potential use cases suggested by the community include:

  • Monitoring server error rates.
  • Analyzing security footage or environmental data.
  • Serving as a low-cost, always-on inference engine for home automation.

Regarding the feasibility of such setups, one commenter noted:

"It's not the worst thing in the world if you have one lying around. 2bit isn't even that terrible - it's ~80% the accuracy of 27b at full quant... Any long running task you can now just slap a PS5 on it."

Implementation and Requirements

To run PS5LM, users must have a jailbroken PS5 and install specific payloads, including kstuff-lite for kernel patches, etaHEN for homebrew support, and ShadowMountPlus to manage the application on the home screen. The project is built using ps5-homebrew-ui and llama.cpp with ggml-vulkan support.

Due to the nature of the project, some community members have expressed concerns regarding its long-term availability on platforms like GitHub, suggesting that developers should consider hosting projects on decentralized or independent infrastructure to avoid potential corporate takedowns.

Sources

Related

  • Dispatch
    llama.cpp and llama.app: Local LLM Inference and Ecosystem

    llama.cpp and its official home llama.app provide high-performance, hardware-agnostic local LLM inference, now featuring a simplified 'llama serve' command and integration with local coding agents like Pi.

  • Dispatch
    vLLM Day 0 Support for Muse Spark Open: 816B-A42B MoE, DSpark Speculative Decoding, and LoRA RL

    vLLM shipped day 0 support for Meta Superintelligence Labs' Muse Spark Open, an 816B-parameter sparse mixture-of-experts model with 42B active parameters, a 1M-token context window, native image/video/audio perception, DSpark speculative decoding, and LoRA RL rollout support.

  • Project
    OpenBMB/MiniCPM-V-Apps

    OpenBMB’s MiniCPM‑V‑Apps repo provides three on‑device demo apps (iOS, Android, HarmonyOS) that run MiniCPM‑V multimodal LLMs locally via a fork of llama.cpp. It includes build scripts, model‑download instructions, and hardware requirements for models ranging from 1 B to a 35 B MoE that streams experts from SSD.

  • Project
    sophgo/LLM-TPU

    LLM-TPU is an open‑source toolkit that compiles HuggingFace LLM/VLM checkpoints into SOPHGO TPU‑specific bmodels and provides ready‑to‑run Python/C++ demos, supporting quantization, KV‑cache, dynamic shapes, and multi‑chip parallelism.

  • Project
    hogeheer499-commits/strix-halo-guide

    A detailed, evidence‑driven guide for installing and benchmarking local LLMs on AMD Strix Halo / Ryzen AI MAX+ 395 hardware, with scripts, BIOS/kernel tweaks, Ollama/llama.cpp setup, and reproducible performance data.