PS5LM Plus runs llama.cpp on jailbroken PS5 with GPU

PS5LM Plus exists because of a technical constraint the developer ran into. Loading llama.cpp as a payload after jailbreaking limits the program to about 6GiB of memory and leaves the GPU unusable. Building the app to launch from the home screen instead raises the available memory to 12GiB and opens GPU acceleration through Vulkan, combining llama.cpp's Vulkan support with a driver ported for the PS5.

The developer's measurements show what that trade-off buys. A quantized version of the 27-billion-parameter Qwen3.8-27B model generates text at 20.7 tokens per second, while a much smaller 1.9GiB Llama 3.2 3B Instruct Q4_K_M model loaded from a USB drive reached 124 tokens per second. The app also handles model selection and memory allocation automatically, checking free memory and adjusting context and cache settings so users do not have to tune them manually.

Beyond chat, the app's API compatibility with OpenAI and Anthropic means coding assistants can drive the PS5-hosted model. The developer reported that opencode, Codex, and Claude Code each successfully edited a Python file to add a function summing two values. Input processing remains a bottleneck, though: Qwen3.8-27B handles prompts at 35 to 50 tokens per second, so a 20,000-token input can take minutes before a response begins.