affine

What is affine?

affine is Bittensor Subnet 120, a decentralized reinforcement-learning and “reason mining” network that pays miners for producing open-weight reasoning models that can defeat a reigning champion across a set of difficult evaluation environments.

Its practical problem is not generic inference supply, but directed improvement of model reasoning under an adversarial incentive scheme: miners submit model weights, validators test challengers against the incumbent, and rewards concentrate on the model that demonstrably improves performance rather than on the miner that merely serves the most requests.

The subnet’s claimed moat is this tournament-style evaluation loop, described in the Affine GitHub repository, where miners commit a Hugging Face model-and-revision pair on-chain and a challenger must beat the current champion across all configured environments by a margin before it can capture the subnet’s weights. (github.com)

affine is a niche but unusually visible Bittensor application subnet, not a standalone Layer 1 or broad smart-contract platform. As of late June 2026, third-party subnet trackers placed SN120 among the larger Bittensor alpha markets, with SubnetRadar ranking it near the top of the subnet market-cap table and showing a nearly saturated UID set, while Bittensor’s subnet directory framed Affine as “decentralized reasoning model evaluation.”

Those figures should be read as market-structure indicators rather than proof of end-user demand, because Bittensor subnet market capitalization, liquidity and “TVL” are derived from dTAO staking pools and alpha-token exposure rather than from conventional protocol revenue. (subnetradar.com)

Who Founded affine and When?

affine appears to have launched in 2025, with OpenTAO listing SN120 as registered on June 10, 2025, during the post-dTAO phase of Bittensor’s evolution, when individual subnets had become investable alpha-token markets rather than purely validator-directed incentive targets.

Public ecosystem material associates the subnet with the Affine Foundation and Jacob Steeves, known as “Const,” a Bittensor co-founder, although the operating team beyond that is not extensively disclosed in formal corporate-style documentation. This opacity is common in Bittensor subnets but matters for institutional diligence because a subnet’s technical credibility, ownership incentives and governance risk are often concentrated in a small founder or operator group rather than dispersed across a mature public company structure. (opentao.ai)

The narrative has evolved from a broad “reasoning market” concept into a more concrete RL competition for open models. Early descriptions emphasized commoditizing reasoning and coordinating multiple subnet resources; the current codebase is narrower and more measurable, focusing on model submissions, one-shot hotkey commitments, plagiarism controls, daily challenger windows and specific evaluation environments such as SWE-INFINITE, LIVEWEB, NAVWORLD, MEMORY, DISTILL and TERMINAL. The shift is important because it moves affine away from a vague decentralized-AI story and toward a falsifiable benchmark-driven mechanism, though the tradeoff is that its economic value remains dependent on whether those benchmark wins translate into reusable external inference demand. (github.com)

How Does the affine Network Work?

affine is best understood as an application-specific incentive layer inside Bittensor rather than an independent consensus network. It does not use its own PoW, PoS or DAG consensus; settlement, registration, staking, alpha accounting and reward distribution are handled through Bittensor’s Subtensor infrastructure, while subnet-level work allocation is governed by Bittensor’s Yuma Consensus mechanism.

In Yuma Consensus, validators submit rankings or weights for miners, and the on-chain process converts those rankings into emissions for miners and validators; in affine’s case, the ranking signal is derived from whether a submitted reasoning model can outperform the incumbent champion under the subnet’s evaluation rules. The Bittensor Yuma Consensus documentation describes this broader mechanism as the algorithm that computes miner and validator emissions from validators’ assessments of miner performance. (docs.learnbittensor.org)

The subnet’s distinctive technical design is its winner-takes-all model-evaluation loop. Miners train or fine-tune models, upload public weights to Hugging Face, and commit a single model revision on-chain; validators then pull the submission, run it against the champion, and only replace the champion if the challenger wins strictly across every active environment.

The Affine FAQ describes one-shot commitments, permanent invalidation for duplicate commits, model-hash plagiarism checks and optional routing of a portion of weights to UID 0 as a burn-like safety mechanism during instability. Validator operations are comparatively lightweight because validators submit weights and monitor backend scoring rather than necessarily running GPU-heavy inference locally, with the validator guide stating that computation is handled by backend services and that validators primarily fetch weights, apply burn configuration and set those weights on-chain. (github.com)

What Are the Tokenomics of sn120?

sn120 is the subnet-specific alpha token for Bittensor netuid 120, with the provided Bittensor explorer reference identifying the asset at subnet address 120. Under Dynamic TAO, every subnet alpha token has its own TAO/alpha pool and a hard cap of 21 million alpha units, mirroring the 21 million TAO cap, while emissions follow a halving-style schedule rather than a fixed fully circulating supply from inception. The Dynamic TAO FAQ and Bittensor subnet documentation explain that staking TAO into a subnet effectively exchanges TAO exposure for that subnet’s alpha token, with the pool reserve ratio setting the alpha price. As of late June 2026, the supplied asset data and third-party dashboards placed sn120 in the low-tens-of-millions market-cap range and around the low-teens dollar price range, but those values are volatile pool outputs rather than stable fundamentals. docs.learnbittensor.org

The utility of sn120 is primarily staking exposure and emissions routing, not payment for gas on an independent chain.

A user who stakes into affine takes exposure to the SN120 alpha token and, through a validator, participates in the subnet’s reward economy; miners seek emissions by producing the winning reasoning model, validators earn by correctly evaluating and setting weights, and stakers earn through the dTAO incentive layer while bearing alpha-price risk.

Bittensor’s emissions system has also changed materially: documentation now states that, as of June 2026, emissions had reverted to a price-based model using subnet EMA token prices, while the flow-based Taoflow model used from November 2025 to June 2026 is deprecated. This matters for sn120 because value accrual depends less on direct fee capture and more on whether staking demand, validator confidence and perceived model utility support the subnet’s alpha market. (docs.learnbittensor.org)

Who Is Using affine?

The observable user base is mostly crypto-native and infrastructure-side: miners, validators, stakers and developers monitoring the model leaderboard. As of late June 2026, SubnetRadar showed affine near full UID utilization, with hundreds of miner slots and a small validator set, which indicates competitive participation for emissions but should not be confused with consumer adoption.

The project’s actual on-chain utility is the production and evaluation of open reasoning models; speculative trading volume in SN120/TAO pools is a separate phenomenon and may dominate short-term market activity even if the underlying inference product has limited external revenue. The public codebase reinforces this distinction by centering miner submissions, validator weight setting and model evaluation rather than a conventional SaaS customer pipeline. (subnetradar.com)

Institutional or enterprise adoption remains limited and should be framed conservatively.

The most concrete integration claim is intra-Bittensor interoperability: OpenTAO and Affine-related materials describe winning models as being deployed through or connected to inference infrastructure such as Chutes, so downstream developers and agent builders can potentially consume the reasoning outputs through API-style access. However, SubnetRadar’s research profile did not show verified 30-day or 90-day external revenue for affine, so there is not yet enough public evidence to treat SN120 as an enterprise revenue asset rather than a promising subnet-level research market. (opentao.ai)

What Are the Risks and Challenges for affine?

Regulatory risk is inherited mainly from Bittensor and TAO rather than from affine alone, but the distinction may not protect subnet alpha holders if U.S. regulators scrutinize the broader network. Grayscale’s Bittensor Trust S-1/A filing states that the trust intends to list on NYSE Arca under the GTAO symbol if its registration and listing process becomes effective, but it also discloses that the trust is not currently staking its TAO and that regulatory approval is not assured. Earlier Grayscale risk disclosure also warned that TAO’s early distribution and the role of the Opentensor Foundation could make the risk of a security classification higher than for Bitcoin-like assets. For sn120, the more immediate centralization risk is operational: a small number of validators, a founder-associated subnet, backend-controlled scoring services, and a winner-takes-all emissions rule can all concentrate influence even if the mining process is nominally permissionless. sec.gov

The competitive threat is two-sided. Within Bittensor, affine competes for emissions, stake and mindshare against other high-profile subnets such as Chutes, Targon, Templar and other model-training, inference and evaluation networks; outside Bittensor, it competes with centralized AI labs and open-source model communities that can improve reasoning models without needing a tokenized incentive layer.

Its mechanism is also economically fragile: if benchmarks are gamed, if model improvements do not generalize, if validators fail to maintain credible scoring, or if stakers rotate into higher-yield subnets, the alpha token can lose emissions support quickly.

Bittensor’s May and June 2026 protocol changes, including the removal of free owner alpha at subnet registration and the Spec 413 protocol-alpha accounting hotfix, show that the dTAO economy is still being actively revised rather than operating as a settled monetary design. (tao.media)

What Is the Future Outlook for affine?

affine’s outlook depends less on price appreciation than on whether it can prove that open, adversarial RL incentives produce durable reasoning improvements that external users actually want to consume.

The verified near-term technical direction is visible in the repository: newer miner guidance requires Qwen3.6-35B-A3B fine-tunes for new submissions, enforces one-shot commits, checks model size and chat-template safety, and relies on daily challenger windows that process one model against the incumbent. At the protocol layer, Bittensor’s June 2026 emissions documentation and Spec 413 upgrade indicate a changing base environment for all subnets, including affine, with emissions now tied again to price-based EMA mechanics and alpha accounting more tightly handled during subnet dissolution.

The structural hurdle is therefore not merely shipping more code; affine must show that its scoring environments resist overfitting, that validator-controlled evaluation remains credible, and that reasoning-model outputs can become useful infrastructure rather than a circular emissions competition. github.com