Table of Contents
Refinery has gone live on Bittensor Subnet 125 as an incentivized optimizer-research market, turning the subnet into a competition over how gradients become parameter updates, one of the lower-level decisions behind language-model training.
Instead of rewarding miners for inference, data delivery, or rented compute, Refinery asks them to submit Python optimizer code. The validator then trains a shared model under a fixed budget and measures whether the submitted update rule produces a lower held-out loss than the current frontier.
The project has published a public site, dashboard, and source repository for the launch. Its first six-week target is deliberately narrow. Refinery wants to find an optimizer that can beat AdamW and Muon on a standardized small language-model run.
That makes Refinery less like an AI services marketplace and more like a research market. The subnet is trying to use Bittensor emissions to coordinate a global search over training algorithms, with rewards tied to confirmed improvements rather than availability or participation.
How Refinery's Optimizer Market Works
Refinery's core mechanism follows a commit, reveal, evaluate, and confirm process.
Miners first publish a SHA-256 hash of their optimizer source code during a commit window and pay a round fee. After commits close, they reveal the source. The revealed code must match the original hash, which prevents participants from changing a submission after seeing what others have entered.
The source then passes through a static gate that is intended to reject code that is not a plain PyTorch update rule. If it qualifies, the validator evaluates the optimizer by training the production model inside a sandbox that cannot reach the network. The held-out loss is recomputed from the saved checkpoint in a clean process that did not run miner code.
A single strong run is not enough to move the frontier. If a submission clears the required bar, the validator funds a second confirmation run. Only if the confirmation also clears the threshold does the result count. Refinery uses the worse loss of the two runs as the new frontier, which makes the mechanism more conservative than simply rewarding the best observed result.
The design is meant to focus the competition on algorithmic improvement. Miners serve source code; the validator performs the training. A slower optimizer does fewer steps inside the same wall-clock budget, while a faster one cannot exceed the scheduled training horizon. The learning-rate schedule belongs to the submission, so miners are responsible for both the update rule and how it is used during the run.
What Refinery Adds To Bittensor
Refinery's thesis is that common optimizers are important milestones with room left to improve on them.
AdamW remains a standard method for language-model training because it decouples weight decay from Adam's adaptive update. Muon has recently drawn attention as another optimizer direction for large-scale training. Refinery is starting with those methods because optimizer design is a broad but measurable target: the model, data, and compute budget can stay fixed while miners explore different rules for using gradients, parameters, accumulated state, and schedules.
The project argues in its research thesis that small, standardized training runs can be a practical way to compare optimizer ideas before attempting larger experiments. Its working hypothesis is that runs costing roughly $100 to $1,000 in compute can reveal relative optimizer performance that may transfer to training budgets 10 to 100 times larger at similar model sizes.
Refinery adds a different kind of subnet experiment to Bittensor.
Many AI subnets focus on serving outputs, coordinating inference, collecting data, routing model work, or supporting application infrastructure. Refinery instead points emissions at open research. Its miners are paid to propose a better part of the training process instead of returning a chat response or hosting a model endpoint.
