Developer abhi245y has released PS5LM Plus on GitHub, an open-source project that turns a jailbroken PlayStation 5 into a local AI server. The project runs llama.cpp as a native app on the console and directly uses the PS5 GPU. Users can browse models from Hugging Face on a TV with a controller, download and load them locally, and let other computers on the same local network call the system through an API.
The project has circulated widely across online communities in recent days, though the BlockTempo report said many reposts left out key conditions such as model quantization and the need for a jailbroken console.
A native app unlocks GPU access and more usable memory
PS5LM Plus is a fork of PS5LM, the original project developed by Mert Cobanov. According to the original project site, the first version ran as a payload on a jailbroken PS5, could use only the CPU, and supported models up to about 1.6 GB.
In the project README, abhi245y said the payload route could access only about 6 GiB of memory and could not use the GPU. Moving to a native app allows access to 12 GiB of direct system memory and the GPU. The inference backend uses ggml's Vulkan backend together with the open-source RADV driver ported to PS5 by community developer Mihawk-99. Model weights and cache share a single 11.44 GiB memory region.
Benchmarks published by the developer
The posted test results show that Qwen3.8-27B in the UD-Q2_K_XL quantized version, sized at 9.2 GiB with a 32k context window, runs at about 20.3 tok/s. The UD-IQ2_XXS version, sized at 6.8 GiB with a 64k context window, runs at about 20.7 tok/s.
Gemma 4 E4B reaches about 61 tok/s. Llama 3.2 3B Instruct Q4_K_M, at 1.9 GiB, reaches about 124 tok/s. For comparison, the same 27B model running on CPU through the payload method produces only about one token per minute.
What the numbers do and do not mean
The report noted that the 27B model fitting on PS5 is a heavily quantized roughly 2-bit version, with lower precision than the 4-bit or 8-bit formats commonly used. BlockTempo also cited a roundup by XenoSpectrum saying the developer's connection documentation lists prompt ingestion speed for the 27B model at only around 35 to 50 token per second. With long prompts, the first response may take several minutes to appear.
The setup also carries a fairly high entry barrier. The project supports only jailbroken PS5 systems running firmware 7.00 to 13.60, and users must preload payloads including kstuff-lite, etaHEN, and ShadowMountPlus. The project itself does not include any Sony SDK, keys, or model weights.
Because jailbreaking allows the console to run software not signed by Sony, the report said this is not a Sony-supported usage mode and may affect warranty coverage and online service rights. Users need to assess that risk on their own.
Why PS5 hardware can deliver usable local inference
According to specifications published on the official PlayStation Blog, the PS5 uses an AMD Zen 2 eight-core CPU, an RDNA 2 GPU, and 16GB of GDDR6 memory shared by the CPU and GPU, with 448GB/s of bandwidth.
When a large language model generates text, each token requires another read of model weights, which makes memory bandwidth a major bottleneck for generation speed. The article said that high-bandwidth unified memory is the main reason a game console launched in 2020 can still deliver usable local inference performance.
More a technical demo than a workstation replacement
The experiment also lines up with a broader move toward local AI on consumer hardware. From high-end Macs to game consoles, developers and enthusiasts are trying to move open models off the cloud and onto local devices.
At the same time, PS5LM Plus remains a hobbyist project that depends on a jailbroken console. The limits from 2-bit quantization and the slow handling of long prompts are hard to ignore. For now, the project looks better suited to technical demonstration than to replacing a dedicated AI workstation.