ChainThink News, March 16: According to official announcements, Bittensor subnet Templar (SN3) has successfully completed the world’s largest decentralized pre-training of a large language model, Covenant-72B, on March 10.

Notably, Covenant-72B is a 72-billion-parameter language model pre-trained by the Templar team on Bittensor Subnet 3, entirely powered by the general internet without reliance on centralized data centers. The model achieved a score of 67.1 on the MMLU (zero-shot) benchmark, outperforming centralized baseline models such as LLaMA-2-70B and LLM360 K2 under identical evaluation conditions. It stands as the largest fully permissionless collaborative language model to date, with over 70 distinct nodes contributing computational resources throughout its training lifecycle. The team has released all model weights and checkpoints under the Apache License.

Following this announcement, Bittensor (TAO) and its subnet tokens surged, with TAO up 54.8% over the past two weeks. Subnet token τemplar rose 194% in the past seven days, now trading at $19.3.