Performances
RD Popularity Index
RD Growth Index
Comparison
Follow Updates
Follow Lists
People
About covenant
Covenant is a decentralized AI training framework. Its core business includes the Templar client for submitting compute tasks, the Basilica protocol that connects independent compute providers, and GRAIL, a decentralized reinforcement learning network. The flagship result is Covenant-72B, a 72-billion-parameter model trained across 20+ independent providers and 70+ individual contributors over 1.2 trillion tokens.
The core narrative is "sovereign AI" — proving that frontier-scale training does not require centralized data centers. Its key technical edge is SparseLoCo and PULSE, which compress gradient updates and only transmit weight deltas, cutting bandwidth needs by over 100x. This makes training on ordinary home internet connections feasible.
The market pain point is twofold. First, centralized AI labs act as gatekeepers, controlling who can build and use frontier models. Second, existing decentralized networks often fail at governance — a few large token holders can dominate decisions. Covenant addresses the first with open infrastructure, and the second with its Conviction Mechanism (BIT-0011), where ownership weight decays over time unless continuously reinforced by locking tokens, making control contestable and aligned with long-term commitment.
In the past six months, the major event was Covenant's public exit from Bittensor on April 9, 2026. Founder Sam Dare called the network's decentralization "a lie," alleging that co-founder Jacob Steeves held effective control and had punished Covenant by halting emissions, revoking channel admin, and unilaterally dismantling its subnet. The announcement caused TAO to drop over 15% in two hours. Since then, Covenant has continued its decentralized AI work outside Bittensor, with the Conviction Mechanism being rolled out on mature subnets including SN3. No major technical releases were announced in this window.
Covenant (formerly Templar) launched decentralized pre-training on Bittensor Subnet 3, using the SparseLoCo optimizer and Gauntlet validation mechanism. After iterating on 1B-8B models, it completed Covenant-72B, a 72B-parameter run with no whitelist and dynamic participation, attracting over 70 independent compute contributors and releasing under Apache License, demonstrating the feasibility of trust-minimized large-scale distributed training.
Due to governance and token-power centralization disputes, the team gradually exited Bittensor, formally leaving in April 2026 to build its own sovereign network (Basilica, Grail, etc.), advancing independent decentralized AI infrastructure.
In the past six months (2026-03-04 to 2026-09-04), Covenant AI made significant strides in decentralized training. The team launched Covenant-72B, a 72B-parameter LLM pre-trained over commodity internet links with permissionless participation, using the SparseLoCo optimizer and Gauntlet reward mechanism. This marked the largest decentralized pre-training run to date, achieving competitive performance against centralized baselines. However, in a major pivot, Covenant AI announced its departure from the Bittensor network, citing governance conflicts with co-founder Jacob Steeves, which caused a 15% drop in TAO token price. Looking ahead, the official roadmap focuses on enabling heterogeneous compute (e.g., mixing data-center GPUs with consumer hardware), scaling node participation, improving training efficiency, and dramatically increasing token counts. The long-term vision includes trillion-parameter decentralized training and pooling smaller compute contributions via mining pools or clusters.
Co-founders include Distributed State (Twitter: https://x.com/DistStateAndMe), 0xNero, and Samuel Dare. Their detailed bios are not publicly disclosed. No other current core executives are identified from available sources.