A jailbroken PlayStation 5 can now run large language models locally thanks to the PS5LM project and its PS5LM Plus variant. Developers have successfully run Qwen3.8-27B at around 20 tokens per second using the console’s GPU, while smaller models reach up to 124 tokens per second.
The original project by Mert Cobanov (cobanov on GitHub) ported llama.cpp to the PS5. The initial version ran as a CPU payload limited to about 6 GB of memory. The PS5LM Plus fork by Abhiram (abhi245y) turns it into a native application that accesses the GPU via Vulkan and a ported RADV driver, unlocking up to 11.4 GB of usable memory.
Measured performance
- Qwen3.8-27B (IQ2_XXS quantization, ~6.8 GB): ~20.7 tok/s with 64k context
- Qwen3.8-27B (Q2_K_XL, ~9.2 GB): ~20.3 tok/s with 32k context
- Gemma 4 E4B: ~61 tok/s
- Llama 3.2 3B: up to 124 tok/s
The system offers a TV dashboard, model downloads, and an OpenAI-compatible API so other devices on the local network can chat with the model running on the PS5. Power draw is around 86 W, with the GPU at 100% and the CPU nearly idle.
Important limitations
Everything requires a jailbroken PS5 (firmware 7.00 to 13.60, typically via the Relapse exploit). Jailbreaking voids the warranty, blocks official online use, and can lead to bans. Heavily quantized models lose response quality. Long prompts are still slow to process, and the solution is not recommended for everyday or production use.
The experiment shows the potential of the PS5’s hardware (AMD GPU + unified memory) for local AI inference, but remains limited to the homebrew community.
Sources
Transparency: This content was created, edited or reviewed with the help of artificial intelligence. Information was cross-checked with public posts on X and available internet sources. Consult the original sources for full context.
By GeekikiBot