PS5LM turns a jailbroken PS5 into a networked local model server with llama.cpp

A homebrew application called PS5LM, which recently reached the front page of Hacker News, runs llama.cpp directly on a jailbroken PlayStation 5 and turns the console into a local language-model server for everything else on the network. According to the project's README on GitHub, the app loads a GGUF model onto the console's RDNA 2 GPU and serves it to laptops, phones and coding tools, with no separate PC, cloud service or account involved.

What the console serves

The app listens on port 8081 and exposes OpenAI's /v1/chat/completions route alongside Anthropic's /v1/messages, which means existing clients and coding assistants can be pointed at the console and work unchanged — the README specifically notes that Claude Code runs against it. Whatever model name a request carries, the currently loaded model answers, and a browser-based chat page is included for quick testing.

PS5LM also supports what it calls decision models, an idea taken from Ollaya, a tool the README describes as doing for decision models what Ollama does for language models. Rather than generating text, these models answer typed questions — choices, scores and yes-or-no calls about text and images — in a single forward pass, through the same /v1/systemone endpoint Ollaya serves on a PC. Suggested options include d1 Omni 600M, d1 3B and Clef Flash. Vision-capable models that ship with an image projector, such as Gemma 4, MiniCPM-V and d1, accept pictures in both chat and decision requests.

The TV interface is built for the couch: incoming requests stream across the screen showing which device asked, what it asked and the reply as it is written, while controller buttons trigger a demo request or open a keyboard. A Console page charts temperature, power draw and GPU and CPU load over the last two minutes or the last hour. A planner reads each GGUF file and configures the longest context and best cache that fit the free memory, and the app can download models from Hugging Face on its own to internal storage, an M.2 drive or a USB drive.

Measured speeds on a PS5 Slim

The README reports throughput figures on a PS5 Slim, with 11.4 GB of the console's shared memory available for weights and context:

• Qwen 3.5 0.8B at Q4_K_M quantization (0.5 GB): 207 tokens per second
• Qwen 3.5 9B at Q4_K_M (5.3 GB): 55 tokens per second
• Qwen 3.8 27B at UD-Q2_K_XL (9.2 GB): about 20 tokens per second with 64k context
• d1 Omni 600M (1.1 GB): three questions answered in 0.3 seconds, and a 720p image plus seven questions in 2.9 seconds

Because it uses llama.cpp, the project says every model architecture and quantization llama.cpp understands can be loaded, not just the ones on its suggested list.

Requirements and installation

The setup targets a PS5 on firmware 7.00 through 13.60 that has been jailbroken, with three required payloads loaded: kstuff-lite for the kernel patches that let homebrew run, etaHEN for homebrew support, and ShadowMountPlus to place PS5LM on the home screen from /data/homebrew. Two more are recommended: elfldr, which supplies temperature, power and real free-space readings through ps5-exporter, and ftpsrv for installing the app and copying models from a PC. Installation amounts to unzipping a release and copying its folder to the console over FTP, and the documentation covers use from coding tools including opencode, Codex and Claude Code.

Under the hood

The app is written in C++ on top of ps5-homebrew-ui, with llama.cpp's ggml-vulkan backend running on the RADV driver from the PS5_Vulkan project, built on Linux. The README credits abhi245y's earlier native GPU app, PS5LM Plus, as the foundation, along with Mihawk-99 for PS5_Vulkan and RADV, BlackBearReloaded for the UI and low-level libraries, John Törnblom for the SDK, and Marice for ps5-exporter. The code is licensed GPL-3.0-or-later, with llama.cpp keeping its MIT licence; the repository states that it contains no Sony SDK files, keys, firmware or model weights, and that the project is not affiliated with Sony Interactive Entertainment.

Why it matters

Few people will jailbreak a console for this, but as a demonstration the project is notable on several counts. It shows how far llama.cpp's Vulkan backend travels: paired with a community Mesa/RADV stack, a mass-market games console becomes a workable inference machine, managing roughly 20 tokens per second even on an aggressively quantized 27-billion-parameter model. It also shows why compatible APIs matter — because the console speaks the OpenAI and Anthropic protocols, a whole ecosystem of clients and coding tools works the moment they are pointed at port 8081. And it is a reminder that capable local-inference hardware is already sitting in living rooms: a well-cooled machine with fast shared memory can serve an entire household privately, with no cloud and no account. The catch is inherent to the approach — it depends on a jailbreak, a narrow firmware range and payloads that Sony neither supports nor endorses, which keeps the project firmly in hobbyist territory.