A new fork of the open-source PS5LM project enables jailbroken PlayStation 5 consoles to run local AI models natively. The software, called PS5LM Plus, uses the console’s GPU and additional system memory, lifting inference speeds well beyond earlier CPU-only attempts.
The original PS5LM project ported llama.cpp to jailbroken PS5 hardware. The new fork takes the concept further by running as a native PS5 application rather than a payload, allowing it to access more of the console’s resources.
The biggest gain comes from GPU access. Reports say the fork can run a 2-bit quantized version of Qwen3.8-27B at more than 20 tokens per second, compared with roughly one token per minute when the same model ran on the PS5’s CPU.
For smaller models, the console can reportedly exceed 100 tokens per second. That makes local inference practical for lightweight chat, coding assistance, and experimentation, although large models still remain slower than modern desktop GPUs. PS5LM Plus also expands usable memory. The native application can access about 12 GiB of memory alongside the GPU, compared with roughly 6 GiB in the earlier CPU-only setup.
That additional headroom allows the console to load larger quantized models. The project has demonstrated Qwen3.8-27B, a 27-billion-parameter model, running locally on the console.
Developer abhi245y built PS5LM Plus on Mert Cobanov’s PS5LM project, which itself ports llama.cpp to the PS5. The fork reportedly uses Mesa’s open-source AMD RADV driver and Vulkan compute support to make GPU inference possible.
The software can serve OpenAI- and Anthropic-compatible APIs. This means compatible local tools and development environments can point to the PS5 as though it were a local AI server.
Also Read: Ooredoo Qatar Launches First 5G+ Network: What Changes?