Jailbroken PS5 Runs Any GGUF Model on Its Own GPU

PS5LM is an inference server for a games console. It loads llama.cpp straight onto the PlayStation 5's graphics hardware and answers HTTP requests locally instead of forwarding them to a hosted provider. A TV-sized front end rides along, drawing incoming requests next to console temperature, power draw and GPU and CPU load.

Vulkan is what unlocks the GPU

The port leans on llama.cpp's Vulkan backend, reaching the RDNA 2 silicon through the RADV driver work around PS5_Vulkan. Model weights and context together get 11.4 GB of the shared memory pool, and that ceiling is the single number every model choice has to fit under. The toolchain builds on Linux and ships to the console as a homebrew package.

Getting there is the hard part. The console has to sit on firmware between 7.00 and 13.60 and be jailbroken, with kstuff-lite supplying kernel patches, etaHEN enabling homebrew and ShadowMountPlus wiring the app into the console's home screen. The author also suggests elfldr on port 9021 and ftpsrv on port 2121 for diagnostics and moving files across.

The throughput numbers are the author's own

All figures below come from the project's own runs on a PS5 Slim. Nobody outside the project has published a replication.

A 27B model landing at 9.2 GB is a function of aggressive quantization, not of spare headroom. Spend that headroom on a long context instead and the same budget disappears somewhere else.

Two API dialects on one box

The server answers OpenAI's /v1/chat/completions and Anthropic's /v1/messages, and the project states that Claude Code talks to it over the latter. Decision models get a separate Ollaya-compatible /v1/systemone route. A browser chat interface sits on port 8081, and models that handle images accept them.

Setup stays deliberately plain: pull the ZIP, FTP the PPSA99575 folder into /data/homebrew, then point the app at GGUF files from the suggested list, internal storage or a USB drive. The project is GPL-3.0-or-later while llama.cpp keeps its MIT terms.

What this report cannot confirm

Exactly one independent source was available here — cobanov's own project page. Every throughput figure, the memory ceiling and the firmware range are self-reported, with no third-party measurement to check them against. The page also publishes no explicit list of limitations, so sustained-load behavior, thermals and long-run stability are simply unaddressed rather than ruled out.

None of this makes the setup a practical path for most readers. A jailbreak means keeping a console on old firmware, and it is not a change you undo casually. The interesting part is the demonstration itself: quantized models now fit on consumer silicon that was never sold as inference hardware.