abhi245y/PS5LM-Plus

llama.cpp on a jailbroken PS5: run any GGUF model natively on the console, starting with Qwen 3.8

What it solves

PS5LM Plus enables running large language models (LLMs) directly on a jailbroken PlayStation 5 using its native GPU. By moving inference from the CPU to the GPU, it achieves significant performance gains, allowing models like Qwen 27B to run at usable speeds (over 20 tokens per second) compared to the extremely slow performance of previous CPU-only implementations.

How it works

It ports the llama.cpp inference engine to the PS5 as a native application. It leverages the console's Vulkan-based RADV driver to utilize the GPU's 12 GiB of unified memory. The app includes a dashboard for monitoring hardware metrics (temperature, fan speed, power usage), a model library for managing GGUF files, and a built-in web server that provides an OpenAI-compatible API and a chat interface accessible from other devices on your local network.

Who it’s for

This project is for users with a jailbroken PlayStation 5 who want to experiment with running local LLMs on console hardware. It is designed for those comfortable with homebrew environments, as it requires specific payloads like kstuff-lite and etaHEN to function.

Highlights

  • GPU Acceleration: Uses the PS5 GPU via Vulkan for high-performance inference.
  • OpenAI-Compatible API: Allows integration with external tools like Aider, Claude Code, and various CLI clients.
  • Integrated Dashboard: Provides real-time monitoring of system thermals, power consumption, and memory usage.
  • Model Management: Includes a built-in downloader and library manager that checks model compatibility and memory requirements.
  • Flexible Deployment: Supports loading models from internal storage, M.2 expansion, or USB drives.

Related

  • Project
    ThinkWatchProject/ThinkWatch-Lite

    ThinkWatch Lite is a cross‑platform desktop gateway that sits between AI coding assistants (Claude Code, Codex, Grok Build, etc.) and the remote AI services they call. It logs every request with cost, masks secrets, blocks dangerous tool calls, lets you switch upstream providers without re‑configuring each client, and can fail‑over automatically. Available for macOS, Windows and Linux, it ships with its own core server and can also run that core on a remote Linux machine.

  • Project
    HotRiceNoodles/CrossLink

    CrossLink is an open‑source Go gateway that presents a single OpenAI/Anthropic‑compatible API while routing requests to any LLM provider (OpenAI, Anthropic, Azure, DeepSeek, Qwen, Ollama, etc.). It handles bidirectional streaming translation, intelligent routing with classified failover, per‑tool RBAC, pluggable guardrails, observability (Prometheus, OpenTelemetry), and includes a Vue‑3 admin dashboard. Deployable via Docker Compose, it supports Redis caching, PostgreSQL storage, and optional Chinese‑national cryptography for air‑gapped environments. Licensed Apache 2.0.

  • Project
    dongri/openai-api-rs

    An unofficial Rust client library that provides convenient access to the OpenAI API and compatible endpoints like OpenRouter.

  • Project
    SuperJJ007/CSSwitch

    A tool that lets Claude Science users connect to third-party model APIs and custom endpoints, providing an isolated environment for switching between different LLMs.