Jailbroken PS5 runs 27B AI model at 20 tok/s

A jailbroken PlayStation 5 runs Alibaba's Qwen 3.8 27B language model at about 20 tokens per second, according to the developer's GitHub page. Wccftech reported the result on the 14th of October 2026.

Tokens are the word pieces a model writes. The work runs on the console's own graphics chip, with no cloud service involved.

Below we break down the speeds on one PS5, by the developer's logs

The tool is PS5LM Plus, a free fork of Mert Cobanov's PS5LM project. A developer called Abhiram publishes it on GitHub.

It runs llama.cpp, a popular engine for local AI models, through the console's graphics chip. Meta's Llama 3.2 3B reaches 124 tokens per second, the GitHub page says.

Google's Gemma 4 E4B reaches 61. Qwen 3.8 27B, shrunk to about 2 bits per value, reaches 20.7.

The same 27B model made about one token a minute when it ran on the console's main processor, the page says. The app gets 12 GiB of memory, double the 6 GiB a simpler payload got.

A jailbreak is required, on firmware 7.00 to 13.60

A jailbreak unlocks the console to run software Sony did not sign. The project's research notes say the exploit covers firmware 7.00 to 13.60, on the PS5 and the PS5 Pro.

Firmware 14.00 closes the hole, the same notes say. Wccftech wrote that jailbreaking voids the warranty and can bring bans from online play.

What is confirmed and what is not

The speeds come from the developer's own logs and have not been tested independently, Back2Gaming wrote on the 15th of October. It named a PS5 Slim as the test console.

No source has published speeds measured on a PS5 Pro. PS5LM Plus says it has no tie to Sony or Alibaba.

Key facts

Qwen 3.8 27B runs at about 20 tokens per second on a jailbroken PS5.

Smaller models reach 61 and 124 tokens per second.

The app uses 12 GiB of the console's 16 GB of memory.

The jailbreak works on firmware 7.00 to 13.60 only.

The code is free on GitHub.