PS5LM Serves Qwen 3.8-27B at ~20 Tokens/Sec on Jailbroken PS5
TL;DR
- Mert Çobanov's PS5LM runs Qwen 3.8-27B at roughly 20 tokens per second on a jailbroken PS5's GPU, versus about one token per minute on its CPU.
- Smaller models hit 207 tok/s (Qwen 3.5 0.8B Q4_K_M) and 55 tok/s (Qwen 3.5 9B Q4_K_M); the 27B runs with a 64K context window.
- The server exposes OpenAI's /v1/chat/completions and Anthropic's /v1/messages, loading GGUF weights into the RDNA 2 GPU via Vulkan with 11.4 GB of shared memory.
On a jailbroken PlayStation 5, Mert Çobanov's PS5LM is pushing Qwen 3.8-27B through the console's GPU at roughly 20 tokens per second. The same model forced onto the PS5's CPU, as one observer summarized it on X, "manages about one token a minute."
The port runs on PS5 firmware 7.00 to 13.60 and loads GGUF weights into the console's RDNA 2 GPU via Vulkan, with 11.4 GB of shared memory for model and context. Smaller models move much faster: 207 tokens/second for Qwen 3.5 0.8B (Q4_K_M), 55 for the 9B version, and the 27B UD-Q2_K_XL variant lands at that ~20 figure with a 64K context window. The README pitches it as "run any GGUF model natively on the console, starting with Qwen 3.8," and the server speaks OpenAI's /v1/chat/completions and Anthropic's /v1/messages so standard clients treat the console as a local host.
Memory bandwidth is why those numbers work at all. The PS5's AMD GPU carries 16 GB of GDDR6 and 448 GB/s of bandwidth, and token generation is bandwidth-bound.
Install is for the hobbyist tier: kstuff-lite, etaHEN and ShadowMountPlus payloads, then an FTP copy of a PPSA99575 folder into /data/homebrew. Our edge-AI tracker has logged 53 on-device stories in the last 90 days, mostly on phones, laptops and purpose-built accelerators; this one lands on retail gaming hardware with an OpenAI-compatible endpoint on the home network.
Originally reported by github.com
Read the original article →Original headline: Open-Source PS5LM Runs GGUF Models Natively on Jailbroken PS5 at 55 Tokens/Sec on Qwen 3.5 9B