Launch of the Qwen3.8 neural network with 27 billion parameters

Developer abhi245y presented the PS5LM Plus project, which allows running language models directly on the PlayStation 5 console's graphics processor. The application uses the GPU's computing power to process requests without contacting cloud servers.

The main Qwen3.8-27B model from Alibaba in 2-bit quantization generates about 20 tokens per second. In quality and performance, it is comparable to flagship models like Anthropic Claude Opus 4.6/4.7, surpassing them in programming tasks and agent work.

Initially, the project was based on a port of the llama.cpp engine by Mert Kobanov, where models ran through the CPU at a speed of about one token per minute. The new version moves calculations to the GPU using Vulkan, significantly speeding up performance.

The application consumes up to 12 GB of video memory and supports several ways to load models: from the internal drive, USB drives, M.2 SSDs, or directly from Hugging Face. Users receive an OpenAI-compatible API for integration with external clients.

Management is carried out via a web interface accessible from any device on the local network. The TV screen displays generation statistics, power consumption, and system temperature.

Lighter models demonstrate high speed: Gemma 4 E4B outputs 61 tokens per second, while Llama 3.2 3B outputs 124 tokens per second.

To run the application, a hacked console with installed etaHEN, kstuff, and ShadowMountPlus components is required. The project source code is available under the GPL-3.0 license.