Jailbroken PS5 Runs Qwen3.8-27B Locally at About 20 Tokens per Second

PS5LM Plus runs Qwen3.8-27B on a jailbroken PlayStation 5 at about 20 tokens per second, turning the console into a local model server that devices on the same network can access.

The open-source app is a fork of PS5LM that runs as a native application, with a chat interface and an API. Its GPU access and 11.44 GiB device memory heap let it run larger models than the earlier payload version, which had no GPU access and about 6 GiB of memory.

The 27B model reached 20.3 tokens per second in a 9.2 GiB configuration with a 32,000-token context. A 6.8 GiB configuration with a 64,000-token context and a different quantization format reached 20.7 tokens per second. Both use roughly 2-bit quantization. For comparison, Llama 3.2 3B Instruct Q4_K_M, with a 1.9 GiB model, reached 124 tokens per second.

Text generation speed does not show how long a prompt takes to process before the response begins. For the 27B model, input processing runs at about 35 to 50 tokens per second; a long prompt can mean waiting several minutes for the first response.

On the earlier payload version, the same 27B model ran on the CPU at roughly one token per minute. The PS5 has 16 GB of GDDR6 memory shared between its CPU and GPU, with bandwidth of 448 GB per second. Qwen3.8-27B also underpins Cloudflare’s self-hostable Clef model.

PS5LM Plus requires a jailbroken PS5 running firmware 7.00 through 13.60, with kstuff-lite, etaHEN and ShadowMountPlus loaded. Jailbreaking allows unsigned software to run, a use Sony does not support, and may affect warranty coverage and access to online services.