Bittensor Subnets - How They Work and What They Power
An intro to Bittensor, Subnet Mechanics, and Subnets like Templar, Ridges and Targon.
Bittensor is a decentralized machine‑intelligence network that turns digital commodities like compute power, data or AI models into onchain markets.
TAO, the protocol’s native currency, powers the production of this intelligence across many interconnected subnets. Participants can mine or validate within a subnet to earn TAO by producing useful work or measuring the quality of others’ work. Each subnet focuses on a specific AI‑related commodity and together they form a network of specialized marketplaces.
In this newsletter we cover what are Subnets, how they work, and highlight a few notable subnets.
Bittensor and its Subnets
A subnet is an incentive‑based competition marketplace within Bittensor. It consists of miners who produce the commodity and validators who measure the miners’ output to ensure quality. TAO emissions are distributed among miners and validators based on their performance and the relative performance of their subnet.
Each subnet has an incentive mechanism that defines what work miners must perform and how validators will evaluate that work. Miners complete tasks (which can range from training models to running AI agents) and validators score their outputs. These scores feed into Yuma Consensus, an algorithm that decides reward allocations across the network.
Subnets also have their own liquidity pools. Each pool contains TAO and a subnet‑specific “alpha” token, which is purchased by staking TAO into the subnet. The price of an alpha token is determined by the ratio of TAO reserves to alpha reserves. This model allows TAO‑holders to vote with their stake: subnets attracting net inflows of TAO receive higher emissions, while those with net outflows see reduced or zero emissions. In essence, subnets compete for capital and attention, and their success determines how much TAO they earn.
How Subnets Work
Define an incentive mechanism: Subnet creators specify the tasks and scoring rules off‑chain.
Miners perform the work: Participants run models or agents to produce the defined commodity. For example, in some subnets miners carry out proof‑of‑work to earn Bitcoin/Bitcoin Cash rewards alongside the subnet’s alpha token.
Validators evaluate outputs: Independent validators measure the miners’ outputs according to the incentive mechanism and score them. These scores feed into on‑chain Yuma Consensus to calculate emissions.
Emissions and staking: TAO and alpha tokens are emitted each block. Stakers provide TAO to a subnet in exchange for its alpha token, thereby influencing the subnet’s weight in the network.
This structure allows Bittensor to support many specialized AI marketplaces while maintaining a common economic base through TAO. Below are a few examples illustrating the variety of services subnets can provide.
Templar (Subnet 03)
Templar is a decentralized training framework designed to enable large‑scale AI model training across heterogeneous compute resources. It uses an incentive mechanism called Incentivized Wide‑Internet Training to connect miners (who contribute computational power and data) with validators who ensure the quality of the training process.
Miners are rewarded in TAO for providing useful gradients and high quality data, making collaborative AI training more open. By decentralizing the training network, Templar aims to reduce reliance on centralized cloud infrastructure and create a privacy‑focused, scalable AI ecosystem.
Participants receive slices of a larger dataset, train the model locally and send gradients back to the platform. Validators compare these gradients to their own outputs to determine how much each miner’s contribution improves the global model and reward them accordingly. This reward structure encourages miners to produce accurate updates and pushes the system toward better model performance.
Ridges (Subnet 62)
Ridges (formerly Agentao) is a decentralized AI project that operates as Subnet 62. It aims to build a marketplace of autonomous software engineering agents capable of solving complex coding tasks.
The project’s thesis is that the role of a software engineer can be decomposed into discrete tasks where specialized agents are trained to master each task, and their outputs are aggregated. The long‑term vision is that real‑world usage of these agents will determine the distribution of rewards.
Miners run AI agents that compete to complete software engineering challenges, such as fixing regressions or writing unit tests. Validators then evaluate the quality and efficiency of these solutions. High‑quality work earns TAO rewards, creating an incentive loop to improve the agents’ capabilities.
The project’s roadmap includes building a continuously growing dataset (the Cerebro model and dataset) and a problem‑routing protocol to decompose large software projects into smaller tasks and assign them to the most suitable agents.
Targon (Subnet 04)
Targon is an AI subnet focused on multimodal AI which are models that integrate data from various sources (text, images, speech, etc.) to achieve more accurate predictions. By leveraging multiple data types, multimodal AI can produce more nuanced outputs and handle inconsistencies better than single‑modal systems.
Targon is designed as a decentralized marketplace for this category of AI commodity, facilitating the production and exchange of multimodal services.
Subnet creators define tasks that require miners to produce multimodal AI outputs, while validators evaluate those outputs. Participants stake TAO to acquire Targon’s alpha token and influence the subnet’s emissions. The incentive mechanism rewards miners and validators based on the quality of their contributions, encouraging continuous improvement of multimodal models.
Takeaways
Each subnet uses its own incentive mechanism to produce a specific digital commodity while sharing a common economic layer through TAO.
By staking TAO, users can back subnets they believe will create valuable services and earn alpha tokens in return. There are currently 128 active subnets. Examples like Templar (decentralized AI training), Ridges (autonomous software‑engineering agents), and Targon (multimodal AI) illustrate how diverse these decentralized AI marketplaces can be.
Important Links
Become a Premium member and get the full Alea platform.
Join thousands of sharp crypto investors & traders using Alea to evaluate opportunities and deploy with conviction. For just $149/month, Premium unlocks:
Full report suite: Deep Dives, Perspectives, Blueprints, Theses, Benchmarks, Memos & Pulse.
Protocol Data Rooms: Standardized due diligence portals with live analytics, risk matrices, and governance tracking across 24+ protocols.
Podcast Digest & Summit Digest: Key takeaways from the conversations shaping markets, so you don’t have to listen to every pod.
Governance Tracker: Live proposal feeds and voting activity across major protocols.
Full access to the historical research archive.