cobanov/ps5lm is a private model server that runs GGUF models with llama.cpp on a jailbroken PS5 and serves them over the network. The latest release is v0.9.1. The published FACTS do not spell out what changed in this version, so the changes cannot be confirmed. The project is written in C++ and is released under the GPL-3.0 license.
The project uses the Vulkan backend on the PS5's RDNA 2 GPU. Vulkan is a software path that connects to a graphics device so that it can carry out the work. The running models can be used from laptops, phones, and coding tools through OpenAI- and Anthropic-compatible APIs and browser chat. It shows requests, responses, and console status on screen and also supports image input. It also offers a `/v1/systemone` API for decision models such as Ollaya. According to the documentation, no PC, cloud service, or account is needed while it runs.
Models and context share the PS5's 11.4GB of shared memory, and the documentation describes a feature that picks the model and cache to fit the memory. The difference visible in the public materials is not a proprietary inference algorithm but the combination of a console port and a model server. The processing speeds in the documentation are the author's own measurements, not independently reproduced results. Since neither the full current code nor the performance has been verified, it is better read as a case of setting up a model server on PS5 than as a performance assessment.