A decentralized AI project just learned the hard way that the crypto community is very good at checking receipts. Quasar, operating as Subnet 24 on the Bittensor network, released what it called a groundbreaking 120 billion-parameter model with a 5 million-token context window. Within hours, independent researchers found that roughly 96-98% of its weights were nearly identical to an existing model from Ant Group.

The fallout was swift. The SN24 alpha token dropped approximately 70-75% in a single trading session.

The model and the mismatch

Quasar, spearheaded by SILX AI, had been building toward this moment for months. The project released Quasar-Preview, a model with around 20 billion total parameters, and Quasar-3B back in April 2026. Both were positioned as stepping stones toward a much larger model trained entirely through decentralized methods on the Bittensor network.

The 120B release was supposed to be the payoff. SILX AI marketed the model as a triumph of decentralized training, complete with promised architectural innovations like something called a “Loop Transformer” and “Engram memory.” These features were meant to solve the memory and coherence problems that plague standard transformer architectures when handling extremely long contexts.

When independent analysts dug into the actual model weights, those novel architectural features were largely absent. Instead, the analysis revealed that the overwhelming majority of the weights, somewhere between 96% and 98%, closely matched Ant Group’s Ling-mini-base-2.0 model.

Why this matters for Bittensor’s ecosystem

Bittensor operates on a fundamentally different premise than most crypto projects. Its network incentivizes decentralized machine intelligence through competitive mining and validation, where different subnets focus on different AI tasks. Subnet 24’s focus has been on developing long-context foundation models, the kind of AI that can process and reason over massive documents without losing the thread.

SILX AI has promoted an ambitious vision for Quasar’s future, including a target of 10 trillion tokens in total training and claims that decentralized training methods can reduce costs by 99.5% compared to conventional centralized approaches.

The token crash and investor fallout

Markets didn’t wait for explanations. The SN24 alpha token’s 70-75% decline in a single session is the kind of drop that typically accompanies fraud allegations or protocol exploits, not model architecture debates.

The gap between what was promised architecturally, including the Loop Transformer and Engram memory, and what was actually delivered in the model weights is too wide to bridge with marketing alone.