Your PS5 was built to render alien landscapes and simulate tire physics. Turns out, the same GPU can also run a large language model without sending a single prompt to the cloud.
That is what PS5LM demonstrates. The open-source project ports llama.cpp to a jailbroken PlayStation 5, routing AI inference through the console’s GPU via Vulkan, the same low-level graphics API used by demanding games on PC and mobile.
The result is a game console functioning as a local AI inference machine.
Performance Details
The headline test pits a 27-billion-parameter model against the PS5’s GPU, with results that vary sharply depending on how the model is compressed and which processor handles the work.
The headline test uses Qwen3.8-27B, a model with approximately 27 billion parameters. Two quantized versions produced roughly 20.3 and 20.7 tokens per second in project benchmarks, depending on model format and context configuration.
Without GPU acceleration, the same model on the PS5’s CPU managed approximately one token per minute, according to reporting by allnight.city. That gap illustrates how heavily the result depends on GPU access.
Smaller models push the numbers considerably higher. PS5LM Plus project documentation reports Gemma 4 E4B at approximately 61 tokens per second and Llama 3.2 3B at up to approximately 124 tokens per second.
Hardware metrics during the larger model test, as reported by allnight.city, showed the GPU running at 100% utilization, system-on-chip power draw at approximately 86 watts, and GPU temperature around 54°C.
A post on X by Vaibhav Sisinty summarized the CPU-to-GPU contrast, roughly one token per minute versus around 20 per second, as the headline performance gap from the benchmark. Sisinty’s direct affiliation with the PS5LM development team was not established; the post reflects third-party commentary on the project’s reported results.
PS5LM Plus reportedly extends the project further, adding a television-oriented dashboard, an on-console model library, and an OpenAI-compatible API that allows other devices on the local network to send requests to the console.
What It Actually Requires
Running AI models on a PS5 demands jailbroken hardware, compatible firmware, and careful management of the console’s limited memory.
This setup is not available on a standard retail PS5. It requires a console running compatible firmware, generally versions 7.00 through 13.60, plus a jailbreak and payload loader.
Memory is the central constraint. The PS5 has 16GB of unified system memory, but the operating system and graphics environment reduce what is available for AI workloads. After quantization, tested model files measured approximately 6.8 GiB and 9.2 GiB; those figures represent file sizes, not total runtime memory, which includes additional overhead from context caches.
Quantization compresses model weights to use fewer bits per value, reducing memory demand and speeding up inference. Output accuracy and reasoning quality can suffer compared with a full-precision version of the same model, though the degree varies by model, method, and task.
Project benchmarks, not independent laboratory results, produced the speeds cited here. Quantization format, context length, thermals, and background system activity can all affect real-world performance.
At 20 tokens per second on a heavily compressed model, the PS5 does not match a modern AI workstation. The console’s memory and software environment remain far more restrictive than a dedicated GPU setup.
Jailbreaking carries real risks: security vulnerabilities, system instability, and voided warranty. Users also risk losing PlayStation Network access and may conflict with platform terms of service or local law depending on jurisdiction.
What the Experiment Actually Shows
The project reveals what the PS5’s GPU can do when homebrew developers reach past the firmware and access the hardware directly.
The PS5’s GPU architecture, designed for parallel graphics computation, is well-suited to the matrix operations AI inference requires. When homebrew developers access low-level GPU interfaces directly, the hardware performs tasks Sony never officially marketed or supported as a local AI platform.
Independent testing of sustained thermals, output quality, and power efficiency across different PS5 hardware revisions will be needed before this moves beyond a technical showcase. The project also operates under a standing condition: a Sony firmware update could close or narrow the currently usable jailbreak window.
