Nearly 97% of Quasar's new 120 billion-parameter AI model weights mirror an existing Ant Group model, causing a dramatic 70-75% plunge in its Bittensor subnet token within hours.
Quasar, backed by SILX AI, unveiled the massive model on Bittensor’s Subnet 24, promising breakthrough decentralized training and new AI techniques like Loop Transformer and Engram memory to handle extremely long text contexts. This release was seen as a major step after earlier versions like Quasar-Preview (20B parameters) and Quasar-3B.
Independent researchers quickly found these claims contradicted by an analysis revealing almost the entire model to be nearly identical to Ant Group’s Ling-mini-base-2.0. The model lacked the advertised novel architecture, raising serious questions about the project's integrity and its decentralized training narrative.
Bittensor’s ecosystem relies on genuine decentralized AI development where subnets compete through mining and validation. Subnet 24’s mission has been to pioneer long-context foundational models capable of grasping extensive documents without losing coherence.
With token value crashing sharply on this controversy, the incident highlights the crypto AI space's fierce scrutiny and the risks tied to ambitious claims without transparent proof. Projects like Quasar that promise revolutionary AI must back them with verifiable originality to maintain community trust.
