User tests 1-bit Hy3 quantization retaining quality at 89 GB

A Reddit user has uploaded a 1-bit quantized version of the Hy3 model to Hugging Face, specifically using the iq1m format. The smallest variant of this checkpoint is approximately 89 GB in size.

• The model is available as AngelSlim/Hy3-GGUF on Hugging Face.
• It utilizes iq1m compression to achieve a file size of 89 GB for the smallest end.
• Testing shows the quantized model retains significant quality and coherence despite the heavy compression.
• Examples include generating HTML flight simulators, SVGs of pandas and pelicans, and images of capybaras.

The post highlights that even as models grow larger, 1-bit quantization can maintain usable performance for local inference.