τemplar (SN3)

Overview

τemplar was an AI training project built on Bittensor’s Subnet 3, also called SN3. Its goal was to bring computers connected over the internet into a shared training process. People who contributed useful computing work could earn rewards through Bittensor. The name “SN3” identifies a place on the network; the token associated with that place is called an alpha token. Bittensor’s network-wide token is TAO. (tplr.ai)

The distinction between the project and the subnet matters. τemplar’s original training protocol is dormant following its Covenant-72B model run. Subnet 3 has since hosted a different AI training system, Teutonic. A reader studying SN3 today is therefore looking at a token whose network role has continued while the training method behind it has changed. (github.com)

History & Team

From shared training to model competition

τemplar began as a way to reward open participation in training large language models. An early technical paper described a live Bittensor run involving a 1.2-billion-parameter model. Participants sent compressed model updates, and validators measured how much those updates helped. The team later reported Covenant-72B, a 72-billion-parameter model trained with contributions from peers connected over the internet. Its research paper says the run processed about 1.1 trillion training tokens. (docs.tplr.ai)

The Covenant-72B paper lists Joel Lidin, Amir Sarfi, Erfan Miahi, Quentin Anthony, Shivam Chauhan, Evangelos Pappas, Benjamin Thérien, Eugene Belilovsky, and Samuel Dare as authors. Their published work describes the model and training approach, while the project’s public code preserves the design of the earlier τemplar protocol. After the Covenant-72B run, that repository marked the original training system as dormant. (arxiv.org)

Subnet 3’s later Teutonic work uses competition between complete model versions rather than τemplar’s shared-update method. Teutonic’s technical report names Jacob Steeves, Krzysztof Warchoł, Hubert Korzeniewski, and Antoni Bombała as its authors. It records a model competition that ran from June 2 through August 10, 2026. These dates show why SN3’s present activity and τemplar’s earlier research need to be understood as different stages in the subnet’s history. (teutonic.ai)

Technology & How It Works

τemplar’s shared-update design

Training a language model involves changing many numerical settings, or parameters, so the model becomes better at predicting text. A training computer calculates an update that suggests how those parameters should change. τemplar divided work among miners, which trained on assigned portions of data. They shared their updates with peers, allowing the group to move a common model forward. Validators checked contributions and posted weights to Bittensor that influenced rewards. (github.com)

Sending full updates for a large model across the internet would require a great deal of data transfer. τemplar reduced that burden through compression. Its documented method transformed updates with a discrete cosine transform, selected important values, and shared the smaller result. Participants then gathered peer updates and applied them to their local model copies. Training took place in coordinated windows so contributors were working from a shared point in the process. (docs.tplr.ai)

Validators judged usefulness by testing whether a miner’s update reduced the model’s prediction error, commonly called loss. This tied rewards to an observed improvement rather than simply to a claim that computing work had been done. Covenant-72B expanded the project’s work on training across distant computers, using a communication-efficient approach that allowed peers to join and leave during the run. (docs.tplr.ai)

How SN3’s later mechanism differs

Teutonic coordinates training through complete model checkpoints. A miner trains a challenger model and submits it for comparison with the current leading model, called the king. The validator evaluates both on the same selected text samples. A challenger becomes the new king when its improvement passes the competition’s test. The accepted checkpoint then provides a starting point for later challengers. In Teutonic’s reported reward design, the current king and as many as four preceding kings share the miner rewards allocated through subnet weights. (github.com)

Tokenomics & Utility

SN3’s alpha token

Bittensor introduced subnet alpha tokens through its Dynamic TAO system in February 2025. Each subnet has its own alpha token and a separate supply limit of 21 million units under Bittensor’s issuance rules. SN3’s alpha belongs to Subnet 3, while TAO serves the wider network. Holding an SN3 alpha position is an economic connection to that subnet; ownership alone does not train a model or submit research work. (bittensor.com)

The token’s incentive role comes from Bittensor’s reward process. Its protocol directs portions of newly issued subnet alpha toward miners, validators and their stakers, and the subnet owner. Miners earn according to the weights set for useful work under the subnet’s operating method. Validators evaluate that work, while stakers can support validators through subnet stake positions. The chain’s standard participant split assigns 41% to miners, 41% to validators and stakers, and 18% to the subnet owner, with further rules governing how the validator portion is shared. (bittensor.com)

TAO and subnet alpha also meet in Bittensor’s built-in pools. Adding TAO stake to a subnet converts TAO into an alpha stake position through its pool; removing that stake reverses the process. This links the token’s exchange and staking functions to Bittensor itself rather than to τemplar’s model-training code. As a result, the alpha token can remain part of Subnet 3’s economics even as the subnet changes its research method. (bittensor.com)

Ecosystem & Use Cases

τemplar’s main use case was collaborative AI research: gathering computing power from different participants to train a shared language model. Its published code and papers make the training and evaluation methods available for study. Covenant-72B is a concrete result of that work, and the project also provided a chat interface for the model. The lasting use of the original τemplar system is in its methods for coordinating and measuring contributions across distant computers. (docs.tplr.ai)

The broader research effort continued beyond that training run. In August 2026, Templar Research described Crucible, a system that joins distributed GPUs for large training jobs and builds on earlier work involving Covenant-72B. Crucible illustrates a use for the group’s training research, while SN3’s Teutonic competition illustrates a different use for Bittensor’s miner-and-validator framework: rewarding independently trained model improvements. Teutonic’s report recorded 2,163 completed model comparisons and 203 accepted new kings during its June–August 2026 study period. (tplr.ai)

Advantages & Challenges

τemplar showed how an open network could coordinate work that is usually done inside a single organization. Its compression methods addressed the cost of sending large updates between computers, and its validator tests gave the network a way to measure contributions. The Covenant-72B run demonstrated that the approach could be applied to a large model with peers taking part over the internet. (docs.tplr.ai)

The same design also had demanding practical needs. Large-model training requires substantial computing equipment, while distant participants must exchange updates and keep their model copies in step. The project’s miner guide for its later training phase listed eight B200 GPUs as a minimum setup. Teutonic changes the coordination problem by comparing complete checkpoints, but its model comparisons still require evaluation work. Its report found that accepted model improvements were concentrated among a smaller group of successful participants, even though many distinct accounts submitted challengers. (github.com)

Where to Buy & Wallets

SN3 alpha is available through Bittensor’s native TAO-to-alpha staking pool. A user can add TAO stake to netuid 3 through a selected validator using Bittensor’s btcli interface. The pool converts the TAO into a Subnet 3 alpha stake position. Bittensor’s wallet tools show both free TAO and stake held across subnets. (bittensor.com)

A Bittensor wallet uses a coldkey to hold funds and sign financial actions, and a hotkey as an operational identity for subnet activity. A token holder can use the network’s wallet and staking tools without operating a miner. Miners and validators use hotkeys for their work on the subnet. (preview.bittensor.com)

Regulatory & Compliance

In the United States, the legal treatment of a crypto-asset transaction depends in part on how the asset is offered and the rights or promises attached to it. The U.S. Securities and Exchange Commission’s 2026 guidance explains that a transaction involving a crypto asset can fall under securities law when it meets the conditions for an investment contract. In the European Union, MiCA sets authorization requirements for businesses providing covered crypto-asset services. These frameworks concern the activities and transactions involved, rather than assigning a single legal status to every subnet token by name. (sec.gov)

Shariah compliance is a separate religious assessment. For SN3 alpha, such an assessment would need to consider the token’s staking and reward arrangements, including how TAO is exchanged for subnet alpha and how rewards reach participants. τemplar’s published technical materials describe those mechanisms and its AI research; they do not establish a project-specific Shariah certification. (bittensor.com)

Future Outlook

τemplar’s research remains relevant to the problem it set out to address: training large AI models with computers spread across different locations. Its original Bittensor protocol is dormant, while the team’s later Crucible work applies ideas from distributed training to a separate platform. On SN3, Teutonic has explored another path, letting miners test independent training choices against a shared leading model. The subnet’s history shows that an AI training goal can support more than one way of measuring and rewarding progress. (github.com)

Summary

τemplar brought reward-based, collaborative language-model training to Bittensor Subnet 3 and produced the Covenant-72B research run. SN3’s alpha token belongs to Bittensor’s subnet economy, where TAO staking, validator weights, and miner rewards connect token ownership with network activity. The original τemplar protocol and the later Teutonic competition use different training methods, making their shared subnet a useful example of how decentralized AI projects can evolve. (arxiv.org)