Bittensor's Templar LLM milestone
Bittensor's subnet Templar completed what was billed as the largest decentralized LLM pre‑training—Covenant‑72B—driving a TAO token pop of roughly 54% and surges in subnet tokens, shared on social. The event is being framed as proof that scalable, decentralized AI infra can move token markets.
Covenant‑72B is documented as a 72‑billion‑parameter model. (arxiv.org) The pre‑training run consumed roughly 1.1 trillion tokens and involved “more than 70” distributed participant nodes, with the process recorded on March 10, 2026. (tao.media) Covenant‑72B achieved a zero‑shot MMLU score of 67.1, a result reported as out‑performing LLaMA‑2‑70B and LLM360 K2, and its weights/checkpoints were published under an Apache license. (kucoin.com) The training coordination layer ran on Bittensor’s on‑chain Gauntlet mechanism for Subnet 3, enabling permissionless peer coordination rather than a centralized orchestration service. (arxiv.org) Market metrics showed TAO touched an intraday peak of $293.8 during the post‑announcement run and recorded a roughly 56% gain over a seven‑day span. (coincentral.com) Subnet economic plumbing requires TAO liquidity interaction—subnet tokens are exchanged against TAO in subnet pools—and launching a subnet has historically required locking on the order of ~1,300 TAO, mechanics that concentrate buying pressure on TAO when a subnet like Templar gains traction. (taostats.io) The Templar subnet’s native token τemplar reached about $19.30 and rallied roughly 194% in a seven‑day window following the run, reflecting speculative capital rotating into subnet alpha tokens. (kucoin.com) A parallel institutional catalyst was Grayscale’s Bittensor trust achieving SEC‑reporting status on March 14, 2026, and exchange listings (reported as Upbit) added tradable liquidity that coincided with the model release‑driven flows. (computing.net)
