PS5 Modded Into AI Server: Runs 27B Models, Connects to Codex & Claude Code

What if your PS5 could do more than just play games? A developer going by the handle abhi245y has done exactly that by turning a jailbroken PS5 into a local AI server capable of running large language models.

Building on Mert Cobanov's PS5LM project—a port of llama.cpp—abhi245y created PS5LM Plus, a native GPU application that lets the console run models like Qwen3.8-27B locally. The PS5's 16GB GDDR6 unified memory, with 448GB/s bandwidth, is shared between CPU and GPU. Early versions of PS5LM ran models on the CPU, but that was painfully slow—about one token per minute for a 27B model. The breakthrough came by porting Mesa's open-source RADV AMD driver, enabling Vulkan compute so llama.cpp could run natively on the PS5's GPU. This increased available memory from 6GB to 11.4GB, enough to run a heavily quantized Qwen3.8-27B.

Quantization is key here. At 16-bit precision, a 27B model would need 54GB of memory—impossible for the PS5. Instead, the PS5LM Plus uses the Qwen3.8-27B-UD-IQ2_XXS version, which stores each weight at roughly 2 bits, shrinking the model to just 6.8GB. That allows it to fit and run at an impressive 20.9 tokens per second, with GPU usage at 100% and CPU at just 2%. Smaller models fly even faster: Gemma 4 E4B hits around 61 tokens per second, and Llama 3.2 3B reaches 124 tokens per second. The whole system draws about 86W and stays at a cool 54°C.

But it's not just about raw speed. PS5LM Plus includes a web chat interface accessible on port 8081 over your local network, so you can chat with the model from any device on the same network. It's also compatible with OpenAI and Anthropic APIs, meaning tools like opencode, Codex, and Claude Code can connect to the model running on your PS5. That's a huge deal for developers who want a local AI server without building a dedicated PC.

Of course, there are caveats. The 16GB memory limit means you're stuck with heavily quantized 27B models or much smaller ones, so quality might not match what you'd get from a full-sized model on proper hardware. And to run native llama.cpp on PS5, you must jailbreak your console—which risks getting your PSN account banned if you play online games.

Still, for a machine designed for gaming, becoming a local AI server is a pretty wild achievement. It shows just how flexible the PS5's hardware can be when pushed by creative developers.

Would you risk jailbreaking your PS5 to run AI models? Let us know in the comments!