Gradients
SN56Compete to fine-tune AI models, where the best automated model tuners earn rewards
Gradients went from one-shot AutoML jobs into multi-day tournaments where the validator executes every miner's training script on dedicated infrastructure, then publishes the winning code open source. Continuous on-chain competition over fine-tuning recipes for LLMs and diffusion models.
// Tournaments for AutoML scripts.
Gradients is a Bittensor subnet for AutoML: the practice of letting a system find the best way to fine-tune an AI model for a given task. Instead of one AutoML library trying one strategy, dozens of miners propose their own training scripts and the validator runs them on dedicated GPUs. The best-performing script wins, and at the end of each tournament the winning code is published open source.
The simple version: It is a coding competition where everyone writes their own recipe for fine-tuning the same AI model on the same dataset. A judge actually cooks every recipe, ranks them by how good the final result tastes, and then releases the winner's recipe so the rest of the world can use it.
Centralized equivalent: Google Vertex AI AutoML or Amazon SageMaker Autopilot, except dozens of teams compete on the strategy instead of one company shipping a fixed algorithm, and the strongest strategies are released back into the open.
How it works:
- Miners write AutoML training scripts and submit them to tournaments. Each tournament runs 4-7 days, with new tournaments starting 72 hours after the previous one ends, so the network runs back-to-back competitions. Top performers earn exponentially higher weight, and the first-place script for every tournament is uploaded to github.com/gradients-opensource. The original organic-task path (real user fine-tuning jobs assigned to pools of miners under fixed time limits) continues to run alongside the tournaments.
- Validators execute miners' open-source training scripts on dedicated infrastructure, test the resulting models against secret evaluation datasets, and set weights accordingly. The main validator is run by Rayon Labs, with independent auditors recomputing weights over rolling 7-day windows as a check on the central coordinator.
- The problem it solves: Most AutoML pipelines run one optimizer one way. Real fine-tuning quality depends on choices an ML engineer would spend weeks tuning. Hiring an engineer is $100k+ per year, paying a centralized AutoML service is up to $10k per large fine-tuning run, and either way you only get one strategy.
- The opportunity: AutoML is one of the few AI categories where the output is a recipe rather than a model, and recipes can be measured cleanly. If the tournament format produces strategies that beat the centralized incumbents on the same evaluations, the open-sourced scripts compound into a public asset the community keeps.
- The Bittensor advantage: Competition is the entire point of the protocol. Tournaments amplify it by adding head-to-head rounds, exponentially weighted payouts, and an open-source release at the end of each cycle.
- Traction signals: 920 commits across 7 contributors. Last push earlier today. The past week of work covers tournament tidying, dataset filtering at task creation, training-time controls, and augmentation-strength changes. 16 active miners on-chain. Operated by Rayon Labs, who runs several other Bittensor subnets and has a record of shipping.
Category: Model Fine-Tuning | Centralized Competitor: Google Vertex AI AutoML, Amazon SageMaker Autopilot, Hugging Face AutoTrain
Gradients sits in one of the most commercially obvious positions in Bittensor. AutoML for fine-tuning has clear pricing on centralized platforms (multi-hundred to multi-thousand dollars per job), an obvious customer (any business with proprietary training data), and a measurable output (a tuned model and the script that produced it). The subnet already competes in that market through miner-driven organic fine-tuning. The tournament layer adds a second product on top.
Mechanism:
Two paths run side by side. The original organic-task path takes real user jobs (upload data, pick a model, click start), assigns them to a pool of miners, and gives each miner the same base model and dataset plus a fixed window (3-10 hours for text, 1-2 hours for images) to produce the best fine-tuned model. Validators score against secret test sets the miners never see during training.
The tournament path is newer and more aggressive. The validator runs each miner's open-source training script on its own infrastructure for 4-7 days, ranks them by the resulting model quality, and assigns exponentially higher weight to the top finishers. A new tournament starts 72 hours after the previous one ends, so the cycle is continuous. The first-place script is uploaded to github.com/gradients-opensource at the end of every tournament, which both rewards the winner with attribution and forces ongoing improvement: last cycle's best is now everyone's baseline.
Auditing is the structural counterweight to a single main validator. Independent auditors can download the past 7 days of task results, run their own evaluations, and recompute what the weights should have been. This is closer to verifiable consensus than the typical single-validator subnet design, though it still depends on auditors choosing to participate.
Markets are quieter than development. Price is at 0.02062 TAO, down about 5% over the last month and roughly flat on the week. Market cap is near 106,044 TAO. Pool depth is about 56,647 TAO with a root proportion of 0.16, which means the AMM is mostly organic alpha rather than protocol subsidy. Emission share is currently 0% and net 7-day flow into the subnet is negative, around 576 TAO out. The 60-day low on TaoSwap is roughly the current price, so we are at the bottom of that range.
Development is the strongest signal. 920 commits across 7 contributors, with the top four (wanderingweights, besimray, diagonalge, samoline1) all in triple digits. The last day of commits covers tournament cleanup (dead code removal, terminal-task management), dataset filtering to drop degenerate cases at task-creation time, augmentation-probability and intensity increases, and reduced instruction-training time. This is operational tuning of a live product, not greenfield work.
- Single main validator: Rayon Labs operates the primary validator. The auditor mechanism reduces but does not eliminate the structural risk of one party coordinating most of the network.
- Concentrated stake distribution: Gini coefficient of 0.703 on the top 100 cohort is on the higher side. Large positions could move price meaningfully on exit.
- Crowded category: Multiple Bittensor subnets pursue model fine-tuning and AutoML. Sustained differentiation depends on tournament outputs continuing to beat alternatives, both inside and outside Bittensor.
- Market drift: Price near the 60-day low and net 7-day flow negative. Until tournament results translate into visible product or developer adoption, demand may stay soft.
- Open-source release as double edge: Publishing the winning script after each tournament is a credible commitment to progress and a public good, but it also hands every other miner and every centralized competitor the latest best recipe for free.
Into the next one.
