Templar Covenant-72B

Templar is a Bittensor subnet that ran a 72-billion-parameter language model pretraining from scratch across an incentivized decentralized network. The training run produced Covenant-72B in early 2026. Apache 2.0. The architecture combines on-chain coordination (Bittensor subnet for incentive distribution and validator scoring) with off-chain bandwidth-efficient training (using techniques in the DisTrO / DiLoCo family). Miners contribute GPU cycles, validators score contributions, TAO emissions reward both. Templar matters because it is the most ambitious distributed pretraining run to date by parameter count. Compared to siblings: INTELLECT-1 (10B, single-org coordination via Prime Intellect), INTELLECT-2 (32B, similar coordination), Pluralis (different topology, earlier stage). Templar's distinctive angle is the combination of token-incentivized contributor network (Bittensor) with frontier-ish parameter scale (72B). The 72B size matters because it is roughly where Llama 3 70B sits, so the comparison to centralized frontier-class open weights is direct. Production-readiness: research-grade. The training run completed and the model is published; whether Covenant-72B competes with Llama 3 70B or Qwen 72B on capability benchmarks is the operative question, and the published results suggest it reaches a credible-but-behind level. Strategic significance: this is the existence proof that token-incentivized decentralized pretraining can produce a frontier-class-size open model, not just a research toy. The honest limit: an order of magnitude behind centralized frontier training on time-to-result, and dependent on TAO economics continuing to reward miners adequately.