Z1T
Z1T is a family of sparse transformer-inspired language models built by Extropic to run on its Z1 probabilistic computing chip. Using gated convolutional attention instead of standard softmax attention, Z1T claims over 100x energy efficiency gains versus GPUs, with throughput around 17,000 tokens per second compared to roughly 1,425 for an H100 baseline. Weights and training code are open source.
- Type
- Text
- Weights
- Open
More from Extropic
3
About Extropic
1
Extropic builds thermodynamic computing hardware for AI. The company develops energy-efficient probabilistic computing chips and systems that use out-of-equilibrium thermodynamics to accelerate AI workloads with radically lower energy consumption than GPUs.