affine (SN120)

Overview

An AI competition on Bittensor

Affine, known as SN120, is a Bittensor subnet that rewards improvements to open-weight language models. Its main activity is a continuing competition: a developer submits a trained model, and Affine compares it with the model currently holding the top position, called the king. A successful challenger becomes the new king. Its model weights are then published for others to study and use. (affine.io)

The subnet has its own Bittensor alpha token. This token belongs to subnet 120, while TAO is the currency of the wider Bittensor network. People can obtain the subnet token through Bittensor’s staking and trading system; participants who improve models can receive subnet emissions under Affine’s reward rules. The token and the model competition serve different roles: one supports economic participation, while the other decides which submitted models earn rewards. (docs.learnbittensor.org)

Affine focuses on reasoning that leads to useful actions, especially in software-development tasks. For example, a model may need to reason through a coding problem and produce a command, tool call, or answer in the required format. The project’s central idea is to make better model behavior measurable, then reward developers who achieve it. (affine.io)

History & Team

From a subnet to a public model contest

Affine operates as Bittensor subnet 120 and is run by Affine Tech. Its public GitHub repository, maintained under the AffineFoundation name, contains the validator, evaluation software, and research materials. The project’s published operating records name Jacob Steeves in decisions about competition and payout rules. Together, the public code and validator dashboard show how the contest is organized and how its rules have developed. (affine.io)

A major change came in September 2026, when Affine moved model submissions to a private upload process. Earlier submissions used Hugging Face repository commits. Under the newer process, challengers upload checkpoints to validator-controlled storage; winning checkpoints are published after they earn the crown. That change gave the competition a way to keep losing submissions private while making winning models available for review. (affine.io)

Affine has also revised its scoring and payout rules as the competition has developed. Those revisions matter because a model contest depends on more than a leaderboard: it needs clear submission rules, a repeatable test, and a defined way to assign rewards. (affine.io)

Technology & How It Works

A challenger faces the king

Developers, called miners in Bittensor, train or fine-tune a model checkpoint. Affine’s submission rules specify the model family and the formats a checkpoint must support. A miner registers a subnet identity, uploads the model files through Affine’s private submission process, and signals that the checkpoint is ready. The validator then places it in a queue. Each registered hotkey receives one evaluation slot, so a miner uses a new hotkey for another attempt. (affine.io)

When a challenger reaches the front of the queue, the validator tests it against the reigning king on the same selected tasks. The tasks come from a published collection of agent interactions. A selected test slice is tied to on-chain information, allowing the selection to be reconstructed later. Before the duel, the submitted checkpoint must also pass checks showing that it loads and follows the required chat format. (affine.io)

Measuring reasoning and action

Affine uses a fixed teacher model as a reference for scoring. In simple terms, the evaluator looks at a model’s reasoning and the action that follows it. It checks how the reasoning relates to the task, how well it supports the next action, and how the action fits the reasoning. The challenger and king are scored on the same turns, rather than receiving separate sets of questions. A challenger must clear both a score-margin threshold and other published conditions to take the crown. (affine.io)

This method differs from simply asking another language model which answer it likes best. Affine’s published scoring code defines the comparison, while the public dataset, model weights of crowned winners, and duel records make important results open to examination. An automated review also examines newly crowned models for ways they may have exploited the scoring process; its verdicts are published. (affine.io)

The system combines on-chain and off-chain work. Bittensor records subnet identities and reward weights. Training, model storage, GPU inference, and detailed evaluation happen outside the chain. This split lets Affine test large language models while using Bittensor to distribute subnet rewards. (affine.io)

Tokenomics & Utility

How the subnet token fits in

SN120’s alpha token is part of Bittensor’s Dynamic TAO system. Each subnet has a pool connecting its alpha token with TAO. Staking TAO into a subnet swaps it for that subnet’s alpha, which is credited to a stake position. Removing stake reverses the process through the pool. Alpha held in subnet 120 is therefore distinct from both TAO and the alpha tokens of other subnets. (docs.learnbittensor.org)

The token’s main economic roles are subnet staking and participation in Bittensor’s emissions system. Stakers delegate through validators, while miners compete to earn the rewards assigned to model-building work. Affine’s own rules determine which winning miner hotkeys receive the weights used for its portion of those rewards. (docs.learnbittensor.org)

Affine pays each crown for up to 72 hours after it wins. If several crowns fall within that window, their registered hotkeys receive equal shares of the miner payout weight. A crown can remain the reigning king after its payout window ends, but its earlier win no longer earns under that rule. When there is no eligible crown, the miner emission is burned. This design connects rewards to repeated model improvements, not only to holding the top position indefinitely. (affine.io)

Ecosystem & Use Cases

Tools for builders and researchers

Affine’s most direct users are model developers. They can train a checkpoint, submit it to a shared contest, and see how it performs against the current winner. The published scoring rules and duel records also give researchers material for studying model training, evaluation, and the link between written reasoning and useful actions. (affine.io)

Winning models have a second use beyond the contest. Affine publishes their weights, making it possible for others to inspect or build on them. The project also provides an interface for chatting with the reigning model. For a reader who does not train AI systems, this is the clearest view of what the subnet produces: a model that can be tried, rather than only a token or a score. (affine.io)

The ecosystem includes several kinds of participants. Miners develop checkpoints; the validator runs the tests and sets payout weights; stakers hold subnet alpha through Bittensor’s staking system; and researchers can examine published models and results. These roles connect AI development to a shared reward mechanism. (affine.io)

Advantages & Challenges

What the design does well—and what it must solve

Affine’s strength is its clear competitive target. A new checkpoint faces the current winner on paired tasks, and the scoring procedure is published. Winning weights and detailed evaluation records support follow-up work by other developers. The short payout window also gives successful miners a reason to keep training instead of relying on one past win. (affine.io)

The approach is demanding. Training and testing large models require substantial computing resources, and submitted checkpoints must fit the serving and format rules. Its results also reflect the tasks and teacher model chosen for the contest. Progress on that measure may be useful for the kinds of agent work Affine tests, but the project must keep examining whether the measure captures the behavior it wants to reward. (affine.io)

Affine’s published rules identify a single validator as the judge of its duels, with other permitted validators mirroring its weight choices. Public code, model weights, records, and post-crown reviews provide ways to inspect those decisions, while the evaluation process still depends on the validator’s operation. (affine.io)

Where to Buy & Wallets

Access through the Bittensor network

SN120 is available through the buy-and-sell interface on Bittensor.ai’s Affine subnet page. The native Bittensor route is to use TAO with subnet 120’s staking pool, which credits the resulting alpha to a subnet stake position. This is a Bittensor subnet asset, so the relevant destination is a Bittensor wallet and subnet stake position. (bittensor.ai)

Bittensor’s command-line wallet can create keys, display balances, and show stake across subnets. Bittensor also lists wallet options including Crucible Wallet, TAO.com Wallet, Nova Wallet, and Subwallet. Wallet interfaces differ in how they present staking, so the key distinction is whether an action places TAO into subnet 120 or merely holds TAO in the wallet. (preview.bittensor.com)

Miners use an additional hotkey as their working identity in the competition. Affine’s private submission process requires an Ed25519 hotkey created through its submission client. The wallet’s coldkey manages funds, while the registered hotkey identifies the miner whose model enters a duel. (affine.io)

Regulatory & Compliance

Legal and religious considerations

Affine’s published materials describe its model contest, subnet token, staking, and reward rules. They do not present a Shariah certification for SN120. An Islamic-finance assessment would need to consider the token transaction, the staking arrangement, and the source and form of rewards together; the project’s AI purpose alone does not establish a halal status. (affine.io)

Legal treatment depends on the activity and jurisdiction. In the United States, the Securities and Exchange Commission explains that a crypto-asset transaction can fall under securities law when it meets the conditions for an investment contract, including an expectation of profit from others’ essential managerial efforts. In the European Union, MiCA sets authorization requirements for businesses providing covered crypto-asset services. Those frameworks concern how assets are offered and how services are provided; Affine’s technical description alone does not settle every legal question about a particular transaction or provider. (sec.gov)

Future Outlook

Improving the contest

Affine’s development path centers on making its model comparisons more useful and easier to examine. Its published history shows revisions to scoring, task selection, model serving, and the review of winning checkpoints. Each change can shape the behavior miners train for, because the evaluation defines what a successful challenger must demonstrate. (affine.io)

The longer-term value of the subnet will rest on the models that emerge from that process. Open weights let developers examine a winner, while public results let researchers study why it won. Continued participation gives the contest new challengers to test against the king. Affine’s role in Bittensor is thus tied closely to whether its competition keeps producing useful model improvements. (affine.io)

Summary

Affine (SN120) connects a Bittensor alpha token to a competition for open-weight AI models. Miners train challengers, a validator compares them with the reigning model, and eligible winners receive time-limited rewards. Its public models, rules, and evaluation records make it a distinctive example of using crypto incentives to support AI development. (affine.io)