A jailbroken PlayStation 5 can serve Qwen3.8-27B to devices on its local network, generating about 20 tokens a second while taking minutes to process lengthy prompts.
Output speed topped 20 tokens per second in each configuration: the 6.8 GiB setup with a 64,000-token context recorded 20.7, while the 9.2 GiB setup paired with 32,000 tokens recorded 20.3. Both use quantization near two bits.
Those figures begin after the prompt has been read. That step processes 35 to 50 tokens a second, and long inputs can leave users waiting minutes for a response.
PS5LM Plus is a PS5LM offshoot with chat and an application programming interface (API) for devices on the local network. The app uses the console's GPU and an 11.44 GiB heap for device memory, exceeding an earlier payload build's roughly 6 GiB and lack of GPU support. That older build processed the 27B model on the CPU at about one token per minute.
A 1.9 GiB run of Llama 3.2 3B Instruct Q4_K_M produced 124 tokens per second. The PS5 combines 16 GB of GDDR6 shared by its CPU and GPU with 448 GB per second of bandwidth. Cloudflare's self-hostable Clef is also based on Qwen3.8-27B.
Running the app requires a jailbroken PS5 on firmware 7.00 through 13.60, plus kstuff-lite, etaHEN and ShadowMountPlus. Jailbreaking permits unsigned software, a practice Sony does not support, and can put warranty coverage and online-service access at risk.
