Templar

Templar is a decentralized large-scale AI training network focused on collaboratively training and improving a single shared language model through distributed participation. Operating as Subnet 3 within the Bittensor ecosystem, Templar allows many independent contributors to take part in model training without relying on centralized infrastructure, ownership, or control.

Instead of training models in isolated environments, Subnet 3 coordinates contributions from across the network so that improvements accumulate into a shared model over time. This approach reduces concentration of control, increases resilience, and enables open participation in large-scale model development while remaining aligned with Bittensor’s incentive-driven design.

Templar operates on a decentralized training model that coordinates compute and learning contributions across independent participants. As Subnet 3 within the Bittensor ecosystem, miners train the shared model using their own compute and data, producing updates that reflect learned improvements rather than full retraining runs.

Validators evaluate submitted updates by measuring their impact on model performance using standardized evaluation tasks. Only updates that demonstrate measurable improvements are accepted and rewarded through Bittensor incentives, ensuring that progress on Subnet 3 is driven by meaningful learning gains rather than raw compute expenditure.