Bittensor's SN3 Completes Largest-Decentralized LLM Pretraining Run in History

• Bittensor's Templar subnet (SN3) trained Covenant-72B, a 720B-parameter LLM, using 70 decentralized nodes, achieving a MMLU score of 67.1.
• The decentralized approach outperformed centralized models like LLaMA-2-70B and highlights the viability of decentralized infrastructure for AI development.
• The success has driven significant price increases for τemplar tokens (up 194% in seven days) and demonstrated a strategic shift from data transport to core model building within the Bittensor protocol.

Bittensor's Templar subnet (SN3) has achieved a historic milestone in decentralized AI training. On March 10, 2026, the subnet completed the training of Covenant-72B, a 720B-parameter large language model, using a decentralized network of 70 nodes. The model was trained using public internet data and implemented entirely through the Bittensor Subnet 3 network.

The Covenant-72B model achieved a MMLU zero-shot benchmark score of 67.1, surpassing the performance of prominent centralized models like LLaMA-2-70B. This result validates the potential of decentralized infrastructure to produce competitive AI models without relying on large, centralized data centers.

The decentralized training approach has drawn attention from key figures in the AI and investment sectors. Jack Clark, co-founder of Anthropic, and Jason Calacanis, a Silicon Valley investor, have publicly endorsed the project. This marks a significant transition for Bittensor, which previously operated as a data transporter. The success of this run confirms the platform's strategic shift toward core model building and positions it as a key player in the evolving AI landscape.

What happened

Templar subnet (SN3) trained a 720B-parameter model using 70 independent nodes. The Covenant-72B model was developed using public internet data and achieved a high MMLU score of 67.1. This milestone is considered the largest decentralized LLM pre-training run in history.

Why it matters

The achievement demonstrates that decentralized compute resources can be effectively used for AI model training, challenging the dominance of large, centralized providers. The success of Covenant-72B also validates Bittensor's strategic pivot from data acquisition to core model development. The Apache-licensed model is available for reuse, potentially accelerating innovation in the open-source AI community.

What are the risks and challenges?

Despite the success, there are risks and challenges associated with scaling this model. Efficiency gaps exist when compared to centralized training, and token volatility remains a concern for investors. The τemplar token, which saw a 194% increase in seven days, may remain subject to high price fluctuations. Additionally, the long-term sustainability of decentralized training for larger models is yet to be proven, and further experimentation will be needed to address performance and cost challenges.

What does this mean for investors?

Investors are showing strong interest in Bittensor's progress. The τemplar token's significant price increase suggests growing confidence in the platform's ability to deliver results. However, token volatility and the broader challenges of decentralized training should be closely monitored. The Apache-licensed nature of Covenant-72B also presents potential opportunities for developers and researchers, which could enhance the platform's appeal and ecosystem growth.