PS5LM Plus Runs Qwen 3.8 on PS5 GPU

This is a remarkable demonstration of repurposed hardware, though the headline speed depends on aggressive quantization and a tightly constrained setup.

• –The native app expands usable memory far beyond PS5LM’s original CPU payload, enabling models that previously could not fit.
• –The reported 20 tok/s result uses low-bit Qwen quantizations, so response quality and tool-use reliability may lag behind higher-precision deployments.
• –Prompt processing is substantially slower than generation, making long-context workloads less practical despite the advertised 32K–64K context windows.
• –Jailbreak, firmware, driver, and homebrew requirements keep this experimental rather than consumer-ready.
• –The project shows how llama.cpp portability and open Vulkan tooling can turn discarded gaming hardware into viable local inference devices.