SYLIN

Extropic Z1T

by Extropic · Model · Language & code models

Z1T-0 sparse transformer weights released on Hugging Face with full training recipe; no inference provider, no text inference script

Evidence

  1. Sep 5
    JAX training code on GitHub can train the sparse models on a normal GPU as simulation; it does not run the released weights
  2. Sep 5
    Needs about 10x more compute than GPT-2 for equal quality, but each operation uses far less energy, so total energy is claimed lower