70+ independent participants trained a 72-billion-parameter AI model on Bittensor with no central server — and it outperformed Meta’s equivalent. Here’s how.
The Experiment That Changed the Narrative
On March 10, 2026, a Bittensor subnet called Templar announced that over 70 independent participants — with no central server, no corporate coordination, using standard internet hardware — had collaboratively trained a 72-billion-parameter language model called Covenant-72B. It achieved a 67.1 MMLU score, confirmed in a peer-reviewed arXiv paper as the largest decentralized LLM pre-training run on record.
Why This Matters
Training frontier AI normally requires thousands of $40,000 GPUs, massive data centers with specialized networking, millions in compute costs, and a coordinated engineering team. Only OpenAI, Google, Meta, and Anthropic can do it. Covenant-72B was trained under radically different conditions — independent participants, standard internet, crypto-based rewards, no central coordinator.
How Decentralized Training Works
Key innovations: Asynchronous training (participants contribute at their own pace), gradient compression (reducing data transmitted over standard internet), economic quality control (Bittensor’s validation mechanism penalizes bad actors), and fault tolerance (participants join and leave freely).
The Implications
Covenant-72B proves frontier AI training doesn’t require billion-dollar budgets. It validates Bittensor’s economic model for the hardest task in AI. It creates new GPU demand for Nvidia. And it introduces a new form of competition for centralized AI labs — not another startup, but an open protocol anyone can contribute to.
The question is no longer “can decentralized AI training work?” The question is now “how fast can it scale?”
The economics are fascinating. The story is even better.
The TAO Story goes beyond tokenomics — the vision, the people, and the experiments that built the world’s first decentralized AI economy.
Learn More About The TAO Story →