---
title: "Hacked PlayStation 5 Turned Into Local AI Server: 27-Billion Model Delivers 21 Tokens Per Second"
description: "Developer abhi245y has introduced PS5LM Plus, an application that runs LLMs on a hacked PlayStation 5 via GPU. The Qwen3.8-27B model delivers 20.9 tokens/s, while compact models reach up to 124 tokens per second."
date: 2026-10-18T10:18:07.930Z
lang: en
url: https://xab.info/en/posts/hacked-playstation-5-local-ai-server
tags: [playstation-5, ai, llm, jailbreak, vulkan, local-ai, gpu-computing]
author: "Alexander Buran"
publisher: "XAB.info"
---

# Hacked PlayStation 5 Turned Into Local AI Server: 27-Billion Model Delivers 21 Tokens Per Second

![PlayStation 5 console standing next to a server rack with blue lighting](https://xab.info/media/2026/10/18/vzlomannaya-playstation-5-ii-server-llm/vzlomannaya-playstation-5-ii-server-llm-1.webp)

## Key Points

- PS5LM Plus offloads LLM inference to the PlayStation 5 GPU via a ported AMD RADV driver and adapted llama.cpp with Vulkan
- Memory available to models increased from 6 to 11.4 GB; Qwen3.8-27B generates 20.9 tokens/s at 100% GPU utilization
- The project provides a web interface and API on port 8081 with support for OpenAI and Anthropic protocols
- Operation requires a console jailbreak, carrying the risk of device or PSN account bans

A developer operating under the alias abhi245y has introduced the PS5LM Plus project, a software solution that allows running large language models (LLMs) directly on a hacked PlayStation 5 console. The initiative is a logical continuation of the earlier PS5LM project created by AI engineer Mert Cobanov. While the initial implementation relied primarily on the console's central processing unit and processed an extremely limited amount of data, the new version shifts the main computational load to the graphics processor, providing full-fledged real-time response generation.

### Offloading to the GPU: How the Architecture Works

The key technical breakthrough in PS5LM Plus is shifting language model inference from the CPU to the GPU. To achieve this, the developer ported the AMD RADV driver from the open-source Mesa project and adapted the llama.cpp library to work through the Vulkan graphics API. This approach made it possible to leverage the compute cores of the PlayStation 5's graphics processor, which were originally designed for rendering game graphics but possess sufficient parallel throughput for the matrix operations typical of neural network computations. As a result, the RAM available for loading models increased from approximately 6 GB to 11.4 GB, opening up the possibility of running medium and large-class models.

### Benchmarks: From 21 to 124 Tokens Per Second

Testing was conducted at a memory bandwidth of 448 GB/s. The Qwen3.8-27B model, containing 27 billion parameters, demonstrated a generation speed of 20.9 tokens per second. Meanwhile, GPU utilization reached 100%, while the central processor was engaged by only 2%, confirming the successful offloading of the workload to the GPU. More compact models run significantly faster: Gemma 4 E4B delivers around 61 tokens per second, and Llama 3.2 3B reaches up to 124 tokens per second. For local use without cloud infrastructure, such metrics make the console a competitive platform for interactive communication with language models.

### Web Interface, API, and Ecosystem Compatibility

PS5LM Plus includes a built-in web interface through which the user can chat with the loaded model. Additionally, the project provides an API accessible via the local network on port 8081. Support for OpenAI and Anthropic protocols allows third-party development tools—OpenCode, Codex, and Claude Code—to connect to the console, turning the PlayStation 5 into a fully functional local node for working with AI assistants within a home or office network.

### Limitations and Risks

Despite impressive results for a console platform, the system remains limited by 16 GB of RAM, which rules out running the largest models. A fundamental requirement for the application to work is a console jailbreak. Sony does not support such modifications, and their use may result in the banning of the device or PlayStation Network account. Users who decide to experiment with PS5LM Plus should keep in mind that in the event of a breach of terms of use, Sony has the right to terminate the account without compensation.

## Fact-Check & Sources

- [Взломанная PlayStation 5 превратилась в ИИ-сервер: модель на 27 млрд параметров выдаёт 21 токен в секунду](https://www.ixbt.com/news/2026/10/18/vzlomannaia-playstation-5-prevratilas-v-ii-server-model-na-27-mlrd-parametrov-vydaet-21-token-v-sekundu.html) - The primary source contains all key facts: developer name, architecture (RADV, Vulkan, llama.cpp), benchmarks (20.9 / 61 / 124 tokens/s), memory volume (6→11.4 GB), bandwidth (448 GB/s), API on port 8081, limitations (16 GB RAM, jailbreak, PSN ban risk). No discrepancies with the article base were found.

## FAQ

### What is PS5LM Plus and who developed it?

PS5LM Plus is an application that allows running large language models on a hacked PlayStation 5. The developer is abhi245y. The project is based on earlier work on PS5LM created by AI engineer Mert Cobanov.

### What models can be run on PlayStation 5 via PS5LM Plus?

Tests included Qwen3.8-27B (20.9 tokens/s), Gemma 4 E4B (~61 tokens/s), and Llama 3.2 3B (up to 124 tokens/s). The total memory available to models is about 11.4 GB out of the console's 16 GB RAM.

### Is a console jailbreak required to run PS5LM Plus?

Yes, running the application requires a PlayStation 5 jailbreak. Sony does not support such modifications, and their use can lead to the banning of the device or PlayStation Network account.

### Can third-party AI tools be connected to PS5LM Plus?

Yes. The project provides an API over the local network on port 8081 and supports OpenAI and Anthropic protocols, allowing the connection of OpenCode, Codex, and Claude Code.