Jailbroken PS5 runs 27B LLM locally at 20.9 tok/s via homebrew llama.cpp port
A Reddit user turned an unofficial llama.cpp port, dubbed "PS5LM," into a native PS5 app, moving inference from the CPU to the GPU via Mesa's RADV driver over Vulkan and raising usable memory from 6GB to 11.4GB. Constrained by the console's 16GB of GDDR6, it could only run a roughly 2-bit quantized Qwen 27B model (6.8GB of weights, 64K context, 91% memory use) at 20.9 tok/s, while smaller models were faster: Gemma 4 E4B at 61 tok/s and Llama 3.2 3B at 124 tok/s, with total power draw of 86W and a 54°C temperature. The trade-offs are heavier quantization errors and weaker reasoning, plus a jailbreak that voids the warranty and risks online bans.
It shows a consumer game console can roughly handle local LLM inference, though memory and quantization limits keep practical use narrow.