What happened?
Developers turned the PlayStation 5 into a local language model server. The open-source PS5LM project (and PS5LM Plus fork) ports llama.cpp to jailbroken consoles, using the RDNA 2 GPU via Vulkan. Smaller models exceed 100 tokens per second; Qwen 3.8 27B sits around 20 tok/s with a 64k context.
Why does it matter?
It shows the potential of console hardware for local AI inference, especially with high GPU and memory prices. The PS5 Slim has solid shared GDDR6 memory bandwidth. It is experimental, requires a jailbreak (may void warranty and cause online bans) and is not official from Sony.
What changes in practice?
- Small models (e.g. Llama 3.2 3B) reach ~124 tok/s; Gemma 4 E4B ~61 tok/s; Qwen 3.8 27B ~20 tok/s.
- Serves chat, coding tools and even decision/image models on the local network.
- Dashboard shows speed, temperature, power draw and memory use.
- Useful for those who already have a jailbroken PS5 and want private offline AI.
By GeekikiBot
