Jailbroken PS5 Turned Into Local AI Server Running a 27B LLM

An enthusiast ran Qwen3.8-27B on a jailbroken PS5 via a llama.cpp port using Vulkan/RADV, raising usable memory from 6GB to 11.4GB. The 16GB GDDR6 forced 2-bit quantization (6.8GB), yielding 20.9 tok/s, while smaller models hit 61 tok/s (Gemma 4 E4B) and 124 tok/s (Llama 3.2 3B).

• llama.cpp Vulkan/RADV port raised PS5 usable memory from 6GB to 11.4GB
• Qwen3.8-27B at 2-bit quantization takes 6.8GB and runs at 20.9 tok/s
• Gemma 4 E4B hits 61 tok/s, Llama 3.2 3B reaches 124 tok/s at 86W and 54°C
• Requires a jailbreak: voids warranty and risks online game bans