PS5LM turns a jailbroken PlayStation 5 into an OpenAI-compatible LLM server
Llama.cpp runs on the console's GPU through Vulkan, and every device on the home network can talk to it.
On a PS5 Slim, the README reports 207 tokens/s for a 0.8B Qwen model, 55 tokens/s for a 9B, and about 20 tokens/s for a 27B model at 64k context. The console gives the model 11.4 GB of shared memory.
What the project does
PS5LM runs llama.cpp on the PS5's GPU (Vulkan backend) to load and serve GGUF models with no PC, cloud or account. It exposes OpenAI-compatible (/v1/chat/completions) and Anthropic-compatible (/v1/messages) APIs on port 8081, a browser chat page, and a TV interface for chat, model management and console monitoring. It also supports image input for models with projectors and "decision models" that answer typed, probabilistic questions.
It requires a jailbroken PS5 on firmware 7.00 to 13.60 with the kstuff-lite, etaHEN and ShadowMountPlus payloads. Elfldr and ftpsrv are recommended for sensors and file transfer.
| Model | Size | Speed (PS5 Slim) |
|---|---|---|
| Qwen 3.5 0.8B, Q4_K_M | 0.5 GB | 207 tokens/s |
| Qwen 3.5 9B, Q4_K_M | 5.3 GB | 55 tokens/s |
| Qwen 3.8 27B, UD-Q2_K_XL | 9.2 GB | ~20 tokens/s (64k context) |
2-bit and 20 tokens a second: cheap, not smart
The 27B result fits the console's 11.4 GB only at 2-bit quantization. A commenter who has used Qwen 3.8 27B warned about the cost:
"anything moderately involved requires xhigh reasoning and it will spend literally tens of thousands of tokens turning over a problem… you won't hit the same quality using a 2 bit quant like this project."
Jacob Miller, a founder who formerly worked at Figma, Atlassian and Microsoft, argued that is fine for unattended work:
"70 cents a day in power cost means you now get ~2M output tokens per day (~$3.50 worth of usage). It's not the worst thing in the world if you have one lying around."
He suggested monitoring error rates or security footage. Treat the PS5 as a slow batch worker for low-stakes background jobs, not a replacement for a hosted assistant.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
Every layer in a transformer is a memory store, and the capacity math shows why attention wins
Alex Wang derives how many key-value pairs a linear layer, an MLP and self-attention can hold before interference equals signal.