QuasarSN24 Price

Quasar is a family of foundation models designed for processing extended context. The models handle millions of tokens in a single pass, with a reported recall rate of 99.9% across the full context window. Unlike traditional transformer architectures that scale quadratically with context length (doubling the context quadruples compute cost), Quasar uses a continuous-time attention mechanism that scales linearly. This reduces both training and inference costs—the project reports inference overhead at one-tenth of conventional models.

The model was trained on 2 trillion tokens using decentralized compute infrastructure via Bittensor Subnet 24, a network where miners contribute computational resources and compete to produce high-performing model variations. Validators assess models on sequence lengths ranging from 8,000 to 2 million tokens, evaluating retrieval accuracy and inference speed. Quasar is available as open-source weights, allowing developers to deploy it for applications such as full-codebase analysis, narrative understanding across entire book series, or legal document review where maintaining context across hundreds of pages is required. The architecture is optimized to run on consumer hardware despite its extended context capabilities.