AdIllustrious7246, the author of a PS5LM fork, reports running Qwen3.8-27B on a jailbroken PlayStation 5 at about 20 generated tokens per second, using IQ2_XXS quantization and a 64k context. The fork runs the model on the console’s GPU through a native app, rather than the original CPU payload. AdIllustrious7246’s project post
What the reported 20 tokens per second covers
That figure describes generated output, not the whole prompt-and-response process. AdIllustrious7246 separately reports prompt prefill—the processing that happens before the model starts answering—at about 35–50 tokens per second. Long prompts can still take time to process before the first generated token arrives. The developer’s account
The developer attributes the prefill constraints to the PS5’s RADV graphics-driver setup, which lacks cooperative-matrix and accelerated integer-dot-product support. The developer says this makes ggml-vulkan use standard fp16 shaders for matrix multiplication. Prompt batches are kept at -ub 64; according to the developer, a GPU submission lasting more than about 10 seconds can cause the console to drop the device. These are details of the fork author’s reported setup.
The IQ2_XXS configuration is also part of the result, not a footnote. It is a very low-bit quantization, a compressed representation of model weights. Some commenters on the project thread questioned whether the quality trade-off makes a 27B model worthwhile in that format; those comments are opinions, not quality measurements. The project discussion
A native GPU app changes the memory picture
The fork author describes the app as a native GPU application using llama.cpp through Vulkan and a RADV port. The author reports about 11.4 GiB available to this app, compared with about 6 GiB for the original PS5LM CPU payload. That larger reported pool is the memory change behind the fork’s ability to run the 27B configuration. AdIllustrious7246’s project post
Why CPU offload does not add PS5 memory
The developer describes the PS5’s memory as a unified pool shared by the CPU and GPU. In this setup, CPU offload moves model weights within that pool; it does not create more memory. The author says offloaded experts would also run on a handful of Zen 2 CPU cores instead of the GPU. The developer’s explanation
The project's stated software requirements
The fork author lists an already-jailbroken PS5 with etaHEN, kstuff, and ShadowMountPlus. The author recommends elfldr and ftpsrv as well. This is a homebrew project running on modified hardware, not a PlayStation feature from Sony. AdIllustrious7246’s project post
