Bittensor Subnet Completes Largest-Ever Scale LLM Pretraining, DeAI Narrative Regression

Bittensor’s Templar Subnet (SN3) has completed the largest decentralized large language model (LLM) pretraining in history, training Covenant-72B with 720 billion parameters using public internet data according to reports. The model was trained across 70 distinct nodes without relying on centralized data centers, demonstrating the feasibility of large-scale decentralized AI projects.

The model was released under the Apache License, allowing widespread reuse and experimentation. Covenant-72B achieved a zero-shot MMLU benchmark score of 67.1, surpassing centralized baselines like LLaMA-2-70B and LLM360 K2 under identical test conditions. This technical success has coincided with a surge in Bittensor (TAO) and its subnet token (τemplar) prices.

Bittensor (TAO) has surged nearly 40% following major AI advancements, including the release of Covenant-72B. The model was trained on 1.1 trillion tokens using fully decentralized methods, allowing participants with GPUs to freely join and leave the network. The launch of Astrid Arena, a new infrastructure tool, further strengthens the Bittensor ecosystem by streamlining the onboarding of developers and AI agents.

Why Did This Happen?

The success of Covenant-72B has demonstrated the viability of decentralized AI training. The model was trained without a centralized cluster or whitelist, using commodity internet connections and the SparseLoCo technique to compress updates according to analysis. This approach allowed decentralized training at scale, making Bittensor’s platform more attractive to AI developers and investors.

The decentralized nature of the training process attracted significant attention from the AI and cryptocurrency communities. A viral social media post by Templar, one of Bittensor’s most active subnets, further fueled demand for subnet tokens and TAO. This demand has driven TAO’s price up by nearly 40% and τemplar up by 194% in the last seven days according to market data.

How Did Markets React?

Bittensor’s price surge has coincided with rising network activity and broader AI sector momentum. The price of TAO broke through the $200–$210 resistance level, with trading volume increasing by 98% to $292.5 million according to trading data. Investors are now watching key resistance levels at $290–$310 and support at $185–$190 to determine the next phase of TAO’s price movement according to technical analysis.

The success of the Covenant-72B project has attracted attention from traders and analysts who see potential in Bittensor’s AI-driven ecosystem. The release of the model has also been supported by rising accumulation signals, with the Accumulation/Distribution line showing a bullish divergence and RSI approaching overbought territory according to technical indicators. These indicators suggest continued upward momentum if the network sustains its current activity.

What Are Analysts Watching Next?

Analysts are closely monitoring the impact of Covenant-72B on Bittensor’s broader ecosystem. The decentralized training approach has shown that large AI projects can be executed without centralized oversight, potentially disrupting traditional AI development models. This could lead to increased participation in subnets and higher demand for TAO tokens.

Investors are also watching how Bittensor’s partnerships and product innovations affect its market position. The launch of Astrid Arena is expected to lower technical barriers for developers and AI agents, potentially accelerating the growth of the Bittensor network. This could lead to increased computational competition and further validation of the DeAI narrative.

The success of the Covenant-72B project has also drawn attention to Bittensor’s broader AI and crypto narrative. As more investors shift capital toward AI-driven projects, Bittensor’s ability to execute large-scale decentralized AI projects could become a key differentiator. This could influence broader market trends and investor sentiment toward decentralized AI platforms.