Bittensor's Quasar Subnet Claims Decentralized Training of 120B Parameter Model

markjeffrey · x · 2026-07-31

Bittensor's Quasar subnet 24 claims to have successfully trained a 120B parameter language model via decentralized compute, featuring a massive 5 million token context window.

According to the author's investigation, the model appears to be legitimate and will be uploaded to HuggingFace soon.

Original post →

More from Models

Models channel →