Bittensor's Quasar Collapses as 'Trained from Scratch' Model Exposed as 96% Plagiarized

markjeffrey · x · 2026-08-02

Quasar (Bittensor Subnet 24) recently claimed to have decentrally trained a 120B parameter model with a 5 million token context window, sparking community awe.

However, they were soon debunked: developer @constreborn exposed that their 'trained from scratch' model is 96% identical to an existing open-source model, suspected of direct cloning. Influencers who previously vouched for it expressed disappointment.

Original post →

More from Fun

Fun channel →