PS5LM brings llama.cpp to a jailbroken PlayStation 5

PS5LM, a homebrew project by developer Mert Cobanov, has produced native PlayStation 5 builds of llama.cpp and completed an initial CPU inference test. The larger goal is to serve GGUF models from the console, with Qwen 3.8 as the starting model. GPU inference remains on the roadmap.

Cobanov’s design ports the general-purpose llama.cpp runtime, which provides C and C++ inference tools for many open-weight models. GGUF is its format for packaging model weights and metadata. A successful port could support Qwen, Llama, Mistral, Gemma and multiple quantization formats through one codebase, although each architecture and backend operation still requires testing on the PS5.

The first executable reaches the CPU

• Build status: llama-cli, llama-server and llama-bench cross-compile and link for the PS5 without changes to upstream llama.cpp source.
• CPU test: Qwen3.5-0.8B reportedly generated text at 14.2 tokens per second using two CPU threads.
• GPU status: The repository contains no completed PS5 GPU inference benchmark.
• 27B status: Qwen3.5-27B has not yet run on the console.

The 0.8B result validates the executable, model loader and basic CPU inference path. It does not substantiate social-media claims of roughly 20 tokens per second for the 27B model on the GPU.

Six gigabytes sets the model ceiling

The PS5 contains 16 GB of GDDR6 shared by the CPU, GPU and operating system. According to the project, a payload process currently has roughly 6 GiB of usable headroom. That allocation is the immediate constraint, regardless of the console’s total physical memory.