The video explores the surprising capability of the PlayStation 5 (PS5) to run large open AI models locally, specifically a 27 billion parameter model called Quinn 3.827B. This model runs on a jailbroken PS5 using a compressed 2-bit version of the model and software called PS5 LM, a port of the popular llama.cpp AI inference tool. The PS5 achieves a text generation speed of about 20 tokens per second, which is impressive given its hardware and compares favorably to mid-range GPUs like the RTX 4060 in terms of memory bandwidth. This bandwidth, at 448 GB per second, is the key factor enabling the PS5 to perform well in generating AI text, as the bottleneck in this phase is memory speed rather than raw computation.
However, the video highlights a critical limitation: the PS5’s AI model reading speed, or “preill” phase, is much slower, at only 35 to 50 tokens per second. This phase involves processing the input prompt before generating any output and is computationally intensive. The slow reading speed is due to software and driver limitations, including lack of support for advanced matrix operations on the PS5’s GPU and a system watchdog timer that restricts long-running GPU tasks. This results in long wait times for complex AI tasks, such as coding agents that require processing large context windows, where initial prompt processing can take over 10 minutes.
Comparisons with Nvidia GPUs reveal the gap clearly: an RTX 5060 Ti with the same memory bandwidth can read input at around 618 tokens per second, making it about 15 times faster than the PS5 in this regard. This difference is attributed to Nvidia’s dedicated tensor cores and mature drivers optimized for AI workloads. Interestingly, running Linux on the PS5 instead of the standard PlayStation OS removes many system restrictions, boosting reading speeds to about 200 tokens per second on the same hardware. This demonstrates that the hardware itself is capable, but the PS5’s default environment imposes significant constraints.
The video also discusses practical considerations and trade-offs. The jailbreak required to run these AI models only works on PS5 firmware versions 7.00 to 13.60, but Sony has since released updates that close this vulnerability and may block online services on older firmware. This means using a PS5 for AI work often sacrifices access to PlayStation Network and new games, making it a tough choice for many users. Additionally, the jailbreak does not persist through reboots, and the system can overheat under sustained AI workloads. For those interested in affordable AI hardware, alternatives like the BC250 mining board, which runs Linux natively and offers similar memory bandwidth, may be more practical.
In conclusion, the PS5 running AI models locally is an impressive engineering feat that challenges assumptions about consumer hardware capabilities. While it can generate text at speeds comparable to some GPUs, its slow input processing limits practical use for complex AI tasks. For existing PS5 owners on compatible firmware who do not need online features, it offers a unique, private AI server at no extra hardware cost. However, for new buyers or those needing responsive AI workflows, dedicated GPUs or Linux-based solutions remain superior. The key takeaway is that AI performance depends heavily on both hardware memory bandwidth and software/driver maturity, with the PS5 illustrating how system constraints can shape real-world capabilities.
