AWERX quantizes DeepSeek V4.1 Flash to 1.62 bits per weight on ELX3

Garrett Kinsman @GeKinsmanOriginally posted 2026.10.20

Open original on X

Agentic engineering is all about pushing the limits of local models. Our team member @ViC305 at @AWERXAI has achieved unprecedented quantization performance of the 10/17 release of Deepseek V4.1 flash using his own methods on ELX3 at 1.62 bits per weight.

This post quotes another post: “DeepSeek V4.1 Flash at 1.62bpw on one DGX Spark: 35.5 tok/s decode”