AI Image
A Jailbroken PS5 Got Turned Into a Local AI Server That Can Run A 27B LLM, Smaller Models Can Reach Impressive Speeds Of 124 Tok/s, But There Are Limitations
Wccftech • • technology
Key Points:
- A user successfully ran the 27-billion-parameter AI model Qwen3.8-27B on a PS5 by shifting processing from the CPU to the GPU, using a heavily quantized version to fit the console's 16GB GDDR6 memory limitations.
- The PS5LM project increased available memory for AI tasks from 6GB to 11.4GB by running llama.cpp natively via Vulkan and a RADV open-source AMD driver port, enabling the console to handle AI workloads more efficiently.
- Due to memory constraints, the AI model was quantized to 2 bits per weight, reducing quality and reasoning ability but allowing it to run at 20.9 tokens per second with the GPU maxed out and low CPU usage.
- Smaller AI models ran faster on the PS5, with token generation speeds up to 124 tok/s, while the console maintained a cool temperature of 54°C and a power draw of 86W during AI processing.
- Running llama.cpp natively on the PS5 requires jailbreaking the console, which voids warranties and risks online bans, highlighting significant trade-offs despite the technical achievement.
