Quasar, operating as Subnet 24 on the Bittensor network, has built an open marketplace where independent miners can train and refine AI models through a competitive evaluation system. The project claims a 99.5% reduction in pre-training costs compared to traditional centralized methods. Quasar launched the Quasar-3B architecture in April 2026 and plans a 10-trillion-token decentralized training run, which is over 30 times the data volume used for GPT-3, trained on roughly 300 billion tokens. The project targets context lengths of approximately 2 million tokens and announced a collaboration with Subnet 56 (Gradients) and Subnet 3 for long-context model training. The TAO token, which powers the Bittensor network, could benefit if Quasar demonstrates that decentralized training can produce models competitive with centralized alternatives.
Source: Read the original article
