How Bittensor subnets work: Yuma consensus, tempo weights, and subnet emissions

Bittensor subnets work as self-contained incentive mechanisms where miners serve an AI commodity and validators continuously price that output by submitting weight vectors on a fixed tempo. Subtensor then runs Yuma Consensus on-chain to convert those weights into subnet emissions and participant payouts, with a separate layer deciding how much emission the subnet receives at all.

Key Takeaways

• A Bittensor subnet is mostly an off-chain rulebook (task + scoring) that becomes economically binding only when its validator weights are settled on-chain by Yuma Consensus.
• Inside each tempo (described as 360 blocks), validators query miners via synapses and submit a weight vector, and Subtensor aggregates those vectors with stake-weighted clipping.
• Emissions are split at the block level in the described architecture: 41% to miners, 41% to validators, and 18% to the subnet creator.
• Subnet emissions are a two-layer allocation problem: one layer decides the subnet’s share of network emissions (root weights and, in the post-2025 framing, dTAO-style stake signals), and another layer decides who inside the subnet gets paid.

How Bittensor subnets are structured

A subnet creator ships an incentive mechanism that specifies two things that matter on a screen: what miners must produce, and what validators will measure. That incentive mechanism lives largely off-chain as code and conventions, then gets launched on the Subtensor blockchain so the subnet is discoverable and can be wired into on-chain settlement.

The template view of a bittensor subnet is deliberately concrete. A subnet has miners, validators, a protocol that defines how they talk, and the Bittensor API glue that lets those actors interact with the on-chain settlement engine, yuma consensus. In the reference template, the “what gets sent” and “what gets returned” is defined in template/protocol.py, miner behavior is in neurons/miner.py, and validator behavior is in neurons/validator.py. The validator’s forward pass and reward logic are explicit surfaces too, in template/forward.py and template/reward.py.

That file layout is not just developer ergonomics. It is the boundary line most explainers blur. The chain is not where the AI task is defined, and it is not where most of the scoring logic runs. The chain’s job is to record stake, accept validator weight submissions, and run the aggregation that finalizes payouts. The subnet’s job is to define the commodity and the measurement.

This is why the broader “TAO subnets explained” framing is useful: the economic object is not “a contract,” it is a market design. The on-chain part only makes it enforceable.

Miner and validator workflow inside subnets

A subnet round is an I/O loop with a standardized payload format and a scoring loop that can be as opinionated as the subnet wants. The template describes Bittensor communication as two processes: preparing data for transit by creating and filling synapses, then processing responses received from axons. Validators send requests through a dendrite, miners serve responses through an axon, and the synapse is the request-response container.

The template’s SubnetsAPI formalizes that interface around two abstract methods: prepare_synapse and process_responses. The point is not the exact method signature. The point is that subnets standardize how a validator turns “I want to test miners on X” into a payload, and how it turns “here are miner responses” into something scoreable.

Operationally, the workflow inside a tempo looks like this:

1. Validators generate queries and build synapses. The synapse encodes the subnet-specific task input. 2. Validators send synapses to a set of miners via dendrite calls and receive axon responses. 3. Validators apply subnet-specific forward and reward logic to turn responses into a weight vector, which is closer to a portfolio allocation than a single score. 4. Validators commit that weight vector to Subtensor for the subnet.

That “weight vector” detail is the part that changes how to reason about incentives. Validators are not just grading miners. They are continuously expressing relative preference across miners, and those preferences are what the chain later aggregates. If a subnet’s reward function is poorly designed, adversarial miners do not need to break the chain. They only need to learn what the validator code pays for.

How Yuma Consensus turns scores into rewards

Every tempo ends with a settlement event. The paper’s described mechanism is that at the end of each tempo, defined as 360 blocks, Subtensor collects validators’ weight vectors for each subnet and feeds them into Yuma Consensus. Yuma then applies stake-weighted clipping to produce miner and validator emissions.

The clipping detail is the on-chain enforcement layer. Validators can submit whatever weights they want, but the aggregation step is designed to reduce the impact of outlier weights relative to the stake-weighted benchmark. The result is that payouts are not “what one validator said,” and they are not “raw accuracy.” They are a consensus outcome shaped by stake.

The same paper also pins down the payout plumbing at the block level in the described architecture. Each block’s minted TAO is split 41% to miners, 41% to validators, and 18% to the subnet creator. Delegators participate by staking to validators and sharing in validator emissions in proportion to their stake.

This is where the two-order-book mental model starts to matter. Inside the subnet, miners and validators are competing over the 41/41 split that Yuma allocates based on aggregated weights. Outside the subnet, the subnet itself is competing for how much emission it receives in the first place. Confusing those layers leads to bad attribution. A miner can improve and still earn less if the subnet’s emission share shrinks. A miner can also earn more without improving if the subnet’s emission share rises.

The tao token is the unit that makes this legible, because all of these emissions and stakes are denominated in TAO at the base layer even when subnet-level tokens exist.

How emissions are allocated across subnets

Subnet emissions are not only about who wins inside a subnet. There is a separate decision about how big the subnet’s pie is. In the pre-dTAO description in the paper, subnets’ relative share of network emissions is governed by the root network, described as the 64 largest validators.

That root-layer control is not hypothetical. The BTCLI root commands expose explicit operations to set weights for subnets by netuid, and to boost or slash a subnet’s weight within the root network. The docs show commands for listing root members and for retrieving the current root weights, which is the on-chain visibility traders and builders actually care about when they are trying to explain why a subnet’s rewards changed.

Post-2025 framing complicates the story, and it needs to be named cleanly. Tokenomist describes a model where staking TAO into a subnet is treated as a signal affecting that subnet’s share of emissions, and it states that since Feb 14, 2025 every subnet has its own alpha token. That is the dtao-era intuition: capital flows become the allocation signal, rather than a committee-like root valuation.

The important reconciliation is that both descriptions point to the same structural reality: there is an “across-subnets” allocation lever. In one era it is explicitly commandable via root weights. In another era it is expressed through stake flows into subnets and their alpha-token pools. Either way, anyone trying to understand subnet emissions has to separate:

1. The subnet’s emission share relative to other subnets. 2. The distribution of that share among miners, validators, delegators, and the subnet owner.

That separation is the difference between diagnosing a subnet that is losing network-level allocation versus a miner that is losing validator weight.

Practical implications and key risks

The incentive mechanism is the product. The paper’s architecture bakes in a revenue share for the subnet creator via the 18% subnet owner share, which makes subnet design closer to launching a marketplace with take-rate economics than deploying a neutral public good. Builders who ignore that will misread why certain subnets get maintained aggressively and others stagnate.

The second-order effect is concentration. The paper’s empirical work documents that stake strongly predicts rewards and that both stake and rewards can be highly concentrated across subnets, which creates a persistent question about whether emissions track output quality or capital positioning. Tokenomist flags a concrete example of this gap with the SN28 case, where emissions were attracted through concentrated stake and validator alignment rather than “strong AI output,” and it notes upgrades like Taoflow as attempts to make simple loops harder to sustain.

For miners and validators, tempo is the operational heartbeat because settlement happens on that cadence. The system does not pay for a single great response. It pays for being consistently scoreable by the validator set by the time weights are submitted each 360-block window.

For TAO holders, the key practical distinction is between choosing exposure to the whole system versus choosing a specific subnet’s allocation dynamics. Tokenomist describes Subnet Zero as an option for those who do not want to pick individual subnets, while still receiving proportional emissions across active subnets.

Near the end of any “TAO subnets explained” discussion, the same question should be asked again: is the subnet competing for a bigger slice of network emissions, or are participants just fighting over the same slice inside the subnet? Most confusion comes from mixing those two ledgers.

The Take

I’ve watched people learn Bittensor by memorizing the slogan “miners answer, validators score,” then get blindsided by the two-layer allocation engine. The expensive mistake is treating a validator’s score like a final verdict. On-chain Yuma Consensus settles weight vectors on a tempo with stake-weighted clipping, so the payout is a consensus outcome, not a judge’s ruling.

I’ve also seen builders ship a clever miner task and underinvest in the validator forward and reward logic. That is where the market is defined. Yuma will faithfully pay whatever the mechanism measures, and the root layer or dTAO-style stake signals decide how much emission the subnet is even playing for. If those two “order books” are not separated, every reward chart gets misread.

Sources

• RaoFoundation bittensor-subnet-template (GitHub)
• Lui & Sun (arXiv)
• Opentensor Foundation BTCLI docs
• Tokenomist