Quasar

Training large language models requires enormous compute resources and infrastructure that individual labs or researchers cannot easily scale or share. Traditional centralized approaches create bottlenecks, waste idle capacity, and make it difficult for distributed teams to contribute training work without managing complex infrastructure. Additionally, ensuring training quality and preventing fraud in a distributed setting requires robust verification mechanisms that existing systems lack.

Quasar Training solves this by distributing model training across independent miners who contribute GPU compute, while validators verify that work meets quality standards. Miners download assigned training tasks with cryptographically signed verification, train model fragments, and upload their work. Validators independently verify miner contributions by checking signatures, data integrity, computational claims, and running their own evaluation. Only validator-approved work is merged into the global model state and contributes to rewards. This creates a verifiable, fraud-resistant training network where compute providers compete fairly based on work quality rather than self-reported metrics.

Quasar Training operates on the Bittensor network (subnet 24) and coordinates Quasar model pretraining across distributed participants. Miners earn rewards from accepted live fragment merges—not self-reported speed or loss—while validators publish weights based on verified work quality. The system manages 10+ trillion training tokens through signed manifests, encrypted presigned grants, and a continuous live fragment sync mechanism where the orchestrator pulls updated model fragments from active miners, validators verify each update, and only approved fragments advance the global model state.

Quasar Training is a decentralized model-pretraining system for Bittensor that coordinates independent miners, validators, and a subnet-operated orchestrator to perform distributed training on the Quasar Preview model. Miners train assigned model fragments and upload signed artifacts, validators verify work quality and publish weights, and the orchestrator assigns tasks, merges accepted updates, and releases checkpoints. The system uses signed manifests, encrypted presigned grants, and a live fragment sync mechanism where validators approve miner contributions before they are merged into the global model state.