PS5LM Runs LLMs Natively on Jailbroken PS5: GPU-Powered AI Inference

PS5LM: Running LLMs Natively on a Jailbroken PlayStation 5

The PlayStation 5, a console designed for gaming, is now being repurposed as a local AI inference machine. The open-source project PS5LM ported llama.cpp to jailbroken PS5 consoles, allowing users to run any GGUF-format language model directly on the hardware. This development transforms the console into a powerful, low-cost AI workstation.

The project, created by developer Mert Cobanov, initially focused on CPU-based execution. It successfully ran a Qwen 3.5 0.8B model at 14 tokens per second on two of the console's Zen 2 CPU cores. However, a subsequent fork named PS5LM Plus has pushed the boundaries further by enabling GPU-accelerated inference, achieving significantly higher performance.

From CPU to GPU: A Leap in Performance

The original PS5LM implementation demonstrated the feasibility of running LLMs on the console, but the performance was limited. The GPU-based PS5LM Plus fork, however, has shown remarkable results. It runs the Qwen 3.8 27B model at over 20 tokens per second, a substantial improvement over the CPU-only approach.

The difference between CPU and GPU execution is stark. When streamed from storage and run through the earlier CPU-based setup, Qwen 3.8 27B reportedly managed only around one token per minute. The GPU-accelerated version achieves this speed at 100% GPU utilization while keeping CPU usage low, demonstrating that the workload is efficiently offloaded to the console's graphics processor.

During testing, the PS5 consumed approximately 86W of power and reached temperatures of 54-55°C, showing that the console can handle sustained AI workloads without thermal issues. This efficiency makes it a viable option for local inference.

PS5LM Plus: A Complete AI Server Solution

PS5LM Plus is more than just a performance enhancement. It includes a native application with a comprehensive dashboard that displays generation speed, GPU and CPU usage, memory consumption, power draw, and temperatures. This provides users with real-time monitoring of the AI workload.

The application can load GGUF models from internal storage, USB storage, or an installed M.2 drive, offering flexibility in model management. It also exposes an OpenAI-compatible API over the local network, allowing other computers or phones to interact with the AI model while inference runs on the PS5. This effectively turns the console into a dedicated local AI server.

Technical Implementation and Compatibility

The PS5LM project is built on a solid technical foundation. It uses llama.cpp as a submodule, with minimal patches to support the PS5's unique environment. The project provides compatibility layers for libc functions the console lacks, and includes probe tools to understand the system's memory and threading capabilities.

The project supports a wide range of model architectures and quantizations that llama.cpp supports, including Qwen, Llama, Mistral, and Gemma, from 1-bit to 8-bit quantizations. This broad compatibility is a key advantage over earlier PS5 LLM projects that hand-wrote GPU kernels for specific models.

The build process requires the ps5-payload-dev SDK and a host LLVM. Running the models requires a jailbroken PS5 on firmware 7.00 to 13.60, with an ELF loader listening on port 9021. The project also includes scripts for setting up the SDK, cross-building llama.cpp, and sending payloads to the console.

Performance Benchmarks Across Models

Performance varies significantly depending on the model size. While the 27B Qwen model achieves around 20 tokens per second, smaller models run much faster. Gemma 4 E4B reached approximately 61 tokens per second, while Llama 3.2 3B achieved as high as 124 tokens per second.

• Qwen 3.8 27B: ~20 tokens/s (GPU)
• Qwen 3.5 0.8B: 14 tokens/s (CPU)
• Gemma 4 E4B: ~61 tokens/s
• Llama 3.2 3B: ~124 tokens/s

These numbers highlight the PS5's potential as a high-performance inference device, especially for smaller models that can run at interactive speeds.

Why This Matters for the AI Community

This development is significant for several reasons. First, it democratizes access to local AI inference by repurposing existing gaming hardware. A PS5, which many people already own, can now serve as a private AI server without the need for expensive dedicated hardware.

Second, it showcases the versatility of the PS5's hardware. The console's GPU, based on AMD's RDNA 2 architecture, proves capable of handling complex AI workloads efficiently. This could inspire further optimization and development for other gaming consoles.

Finally, the project demonstrates the power of open-source collaboration. PS5LM builds on the work of the llama.cpp and ps5-payload-dev communities, and its GPL-3.0 license ensures that improvements are shared back. The project is not affiliated with Sony or Alibaba, and contains no proprietary SDK files or model weights.

Getting Started with PS5LM

For those interested in trying PS5LM, the project provides comprehensive documentation. The GitHub repository includes a roadmap, a list of compatible models, porting notes, and research findings. The demo video and interactive chat page can be accessed at ps5lm.cobanov.dev.

Users need a jailbroken PS5 on supported firmware, an ELF loader, and the ability to transfer models via FTP or USB. The project's scripts handle the heavy lifting, from building llama.cpp to uploading and executing the inference tools.

While the project is still in its early stages, with a roadmap for larger models and further optimizations, it represents a significant step forward in the convergence of gaming hardware and AI. As the community continues to refine the implementation, we can expect even more impressive performance and capabilities from this unlikely AI platform.