Gradients allows anyone in the world to train image & text models - intelligence, simplified.
SN 56. Rank 23 of 32 by emission
Gradients
Emission1.4%of the block reward · rank 23 of 32
Participants7 miners · 12 validators
Registration0.0005 TAO
Verified2026-09-09links checked and scores merged
The programtraining · github.com@0791bb6 · read 2026-09-22
The work
Training code: a GitHub repository and commit, served from a small miner endpoint, that validators build in a container and run on their own GPUs to fine-tune a model on a task. The miner supplies no compute, pays 0.4 to 0.7 TAO to enter a weekly tournament, and a winning repository is republished in public.
code · raw.githubusercontent.com · docs/miner.md
Tournament miners submit training code. Validators clone your submitted repository at the commit you provide, build your Docker image, run your training script
How it is scored
Round one is a single group of every entrant playing three instruct tasks; the top two advance to a knockout, whose winner meets the reigning champion in a six-task boss round. Two entries train the same model on the same task; the winner wins more held-out samples, ties inside 0.01 nats counting for neither, mean loss deciding equal counts. GRPO tasks rank on mean score. Duplicates are eliminated in tiers.
code · raw.githubusercontent.com · docs/miner.md
Knockout instruct, DPO and chat tasks are decided per held-out sample, not on mean loss.
How the pay splits
Winner-heavy per pool. Emission divides into text, image and environment pools anchored at 0.35, 0.25 and 0.25, rebalanced by participation with a 0.05 floor and 0.50 cap. Inside a pool the top three ranks are paid by exponential decay with base 0.25, about 76, 19 and 5 percent; a champion whose boss-round margin passes 0.10 has the excess doubled. Every participant gets 0.0001.
code · raw.githubusercontent.com · validator/scoring/constants.py
TOURNAMENT_PAID_RANKS = 3; TOURNAMENT_SIMPLE_DECAY_BASE = 0.25; TOURNAMENT_PARTICIPATION_WEIGHT = 0.0001
Sources disagree docs/miner.md says emissions split between champions and participants and nothing is burned; validator/scoring/weights.py agrees and rescales the three pools to fill the whole pie; validator/scoring/emission_balance.py at the same commit says anything shaved off the 0.50 cap falls through to burn. EMISSION_BURN_HOTKEY is a placeholder for a champion that defended, resolved to the real hotkey before scoring.
Who judges
Twelve validators were active on the 2026-09-09 chain snapshot. Commit-reveal is off. Validators do the work themselves: they build each entry, run the training on validator-managed GPUs, upload the model and score it against its opponent on a held-out set. Completion is manual: the team reads the winning code for exploits before a tournament is marked complete, with a Friday 14:00 UTC deadline.
chain · hyperparams-2026-09-19.json
commit_reveal_weights_enabled False, weights_rate_limit 100; chain-2026-09-09.json active_validators 12
How often it turns
Emission settles every 360 blocks, about 72 minutes. The puzzle turns weekly per pool: environment tournaments start Monday 09:00 UTC, text 11:00, image 13:00, only when no tournament of that type is active, and each must complete by Friday 14:00 UTC. A champion's weight does not decay between tournaments. A new registration is immune from deregistration for 5,000 blocks, about 16 hours.
chain · hyperparams-2026-09-19.json
tempo 360, immunity_period 5000
Who buys
Anyone with a model and a dataset, by the hour, through the site or one API call to api.gradients.io/v1/tasks/create. The rate card, read from SKILL.md because the pricing page loads only with JavaScript, runs $10 to $50 an hour by text model size and $5 an hour for image models. No customer or usage figure is published; the last news post is from November 2025.
code · raw.githubusercontent.com · SKILL.md
<=1B text: $10/hour; <=7B text: $15/hour; <=40B text: $25/hour; >40B text: $50/hour; image: $5/hour
What has been gamed
One entry under many identities: submissions are de-duplicated in three tiers, identical commit, identical source after normalisation, and a Claude pairwise review of deltas, with cosmetic disguises counted as duplicates and eliminated. Hidden datasets, pretrained weights, compiled or obfuscated code are banned. One hotkey per coldkey per tournament. Before a tournament closes the team reads the winning code for exploits, a manual step.
code · raw.githubusercontent.com · docs/miner.md
the team reviews the winning submission's code to check for exploits and confirm it reflects a genuine training improvement
7 of 7 fields rest on the chain or the code. A manifest is the site's reading of the subnet's program, not a review; each field links its source. All programs · What a field means · Correct a field
What it says
The subnet in its own words: the line the index took from its front door, and what the team has said on the record. All 12 quotes in the map.
Their subnet
Gradients describes itself as a tournament and a race among miners, and as the easiest place in the world to train a model.
Current AutoML platforms leave substantial performance untapped. Testing 180 fine-tuning tasks across models from 70M to 70B parameters, we found that HuggingFace AutoTrain, TogetherAI, Databricks, and Google Cloud consistently produce suboptimal configurations.
Gradients, built on the Bittensor network, attacks this problem through competition. Independent miners race to find optimal hyperparameters, earning rewards proportional to their models' performance. This tournament drives exploration of configuration spaces that single-strategy methods never examine.
When miners profit from finding better configurations, they develop novel optimization techniques, share successful strategies within their organizations, and continuously refine their approaches.
The world's best AutoML platform, powered by Subnet 56 on Bittensor.
With just a few clicks, gradients will be training your model for you, 0 AI training knowledge needed. The easiest and best place in the world to train a model
Bittensor, in their words
Bittensor is a marketplace and a decentralised system with economic incentives; the paper generalises the subnet's result to market dynamics, which is closer to Steeves's market language than to the foundation's.
These findings demonstrate that decentralized systems with economic incentives can systematically outperform traditional AutoML, suggesting market dynamics may be key to achieving superior fine-tuning results.
Built on the Bittensor network, it creates a marketplace where independent miners compete to find optimal configurations.
This suggests a fundamental shift in how we approach AutoML: from better algorithms to better incentive structures.
We have built a platform that makes it easier than ever to train AI, powered by the new, revolutionary Bittensor Subnet 56
Frames market marketplace revolution tournament race
What is said about it
By other voices in the map, verbatim and dated.
Think AutoML on Bittensor and you're on the right track but still selling this group way short.
What it shows
Explained from zero by the index, from the subnet's own materials, for someone who has never heard of Bittensor.
SN56 Gradients is a tailoring service for AI models: a customer drops off a model and a dataset, pays by the hour, and gets back a model fitted to their data, cut by whichever tailor won this week's contest.
The commodity, explained from zero
Fine-tuning is taking a model someone else trained and adjusting it on a company's own examples: support tickets, product photos, legal clauses. The result is a checkpoint, a saved model file, delivered to the customer's Hugging Face account. Gradients sells this by the hour. The rate card, read from a file in the repository because the pricing page loads only with JavaScript, runs from $10 an hour for the smallest text models to $50 for the largest, and $5 an hour for image models.
The customer is anyone: "0 AI training knowledge needed". A job starts from a web form or one API call to api.gradients.io/v1/tasks/create with a model, a dataset and hours to complete. The materials name no customers and give no usage figure. The news page's last post is from November 2025; the boldest claim, from February 2025, is that Gradients beat TogetherAI and Google as "the best 0-Click Training Platform in the World".
Why it is on Bittensor at all
The subnet's argument is that competition finds better training recipes than one company's engineers. Miners do not sell GPU time; they submit training code, and the code that wins is published at github.com/gradients-opensource for anyone to copy. The comparison breaks here: a tailor keeps the pattern, while a winning Gradients tailor must hand the pattern to the public, and the shop's own machines, not the tailor's, do the sewing.
How the work gets done
Miners (the tailors) run a small endpoint that tells validators which GitHub repository and commit to enter in each weekly tournament, and pay an entry fee in TAO. Validators (the judges) clone the code, build it in a locked-down container, run it on validator-owned GPUs, upload the model to Hugging Face and score it against the other entries, deciding knockout rounds "per held-out sample, not on mean loss". Within a tournament the top three ranks are paid roughly 76, 19 and 5 percent of that tournament's emission, the subnet's share of newly minted TAO, and a champion whose margin passes a 0.10 threshold has the excess multiplied by 2.0. Before a tournament closes, the team reads the winning code for exploits.
How you would know it works
The API reference at api.gradients.io/docs is live, and a public endpoint, GET /v1/network/status, reports completed jobs without a login. No benchmark table exists on the site; the research page needs an account.
What is missing
The entry fees disagree: the README quotes one set of TAO fees for the three tournament types, while the miner guide and the fees API quote double. The chain description is a slogan with a typo, "Best AutoML plaftorm in the world", and the discord and additional fields hold the literal word "None". The contact info@gradients.io matches the chain but appears only inside a news post. The only support path is Discord; there is no status page or SLA. The miner guide's clone URL still points at the old rayonlabs organisation.
Go deeper
Sources for this explainer
- https://www.gradients.io/
- https://www.gradients.io/news
- https://www.gradients.io/pricing
- https://api.gradients.io/docs
- https://raw.githubusercontent.com/gradients-ai/G.O.D/main/README.md
- https://raw.githubusercontent.com/gradients-ai/G.O.D/main/docs/miner.md
- https://raw.githubusercontent.com/gradients-ai/G.O.D/main/SKILL.md
Metaphor: a tailoring service with a weekly cutting contest. Every claim is drawn from the evidence set or the subnet's own materials; "(inferred)" marks a conclusion rather than a quote. Corrections.
How legible it is
Four audiences, four questions each, rated on what a first-time reader can find in five minutes. The scale measures findability, not quality.
On-chain identity 7/7- subnet_name
- Gradients
- github_repo
- github.com
- subnet_contact
- info@gradients.io
- subnet_url
- gradients.io
- discord
- None dead
- description
- Best AutoML plaftorm in the world
- additional
- None
Stakers and validators●●●●● 3.0
Should I allocate here? · rank 18 of 32 for this audience
Q1What it is●●●●● 4
Gradients allows anyone in the world to train image & text models - intelligence, simplified.
Output is fine tuned models priced per hour on a JS only pricing page, chain says Best AutoML plaftorm
Q2Who it is for●●●●● 2
Gradients Beats TogetherAI, Google to become the best 0-Click Training Platform in the World.
Dated 2/13/2025, a model gallery with no count, no usage or revenue figure, latest post 11/2/2025
Q3How it resists gaming or fails●●●●● 3
Scoring And Weights section: per type pools, top three ranks paid 76/19/5, champion multiplier, nothing burned, winning code reviewed for exploits, written for miners, no failure modes
Q4Identity and documentation●●●●● 3
identity 7 of 7 but discord and additional hold the literal text None, github and url resolve, info@gradients.io appears in a news post, no staker or validator page
Miners●●●●● 4.0
Can I compete, and what wins? · rank 3 of 32 for this audience
Q1What it is●●●●● 5
Tournament miners submit training code. Validators clone your submitted repository at the commit you provide, build your Docker image, run your training script
with response schema and example code
Q2Who it is for●●●●● 3
You do not need to provide tournament compute.
Fees 0.7, 0.4 and 0.6 TAO confirmed by the fees API while the README says 0.35, 0.20 and 0.30, registration cost not stated
Q3How it resists gaming or fails●●●●● 4
prose with numbers: decay base 0.25, roughly 76/19/5 across paid ranks, 0.10 threshold and 2.0 multiplier for champions, no worked example, no immunity or deregistration policy
Q4Identity and documentation●●●●● 4
guide called the single source of truth with checklist, setup and local testing, repo pushed 2026-09-09, no releases or changelog, clone URL still points at the rayonlabs org
Buyers and enterprises●●●●● 3.8
Can I use this today? · rank 8 of 32 for this audience
Q1What it is●●●●● 5
Begin training with a single simple user interface, or programmatically with our API.
Curl POST to api.gradients.io/v1/tasks/create on the front page, Swagger and openapi.json live
Q2Who it is for●●●●● 4
Competitive rates for model training. Pay for what you use, with no hidden fees.
Rates $10 to $50 per hour by model size and $5 for images render only with JS, customer is anyone
Q3How it resists gaming or fails●●●●● 2
Join the Subnet Discord is the only support path, no status page, no SLA, a network status endpoint is listed in SKILL.md
Q4Identity and documentation●●●●● 4
Swagger API reference plus openapi.json, info@gradients.io matches the chain contact but appears only inside a news post, no contact page
Newcomers●●●●● 3.5
What is this and why does it matter? · rank 15 of 32 for this audience
Q1What it is●●●●● 4
With just a few clicks, gradients will be training your model for you, 0 AI training knowledge needed.
Plain and on the front page, the chain description is a slogan with a typo
Q2Who it is for●●●●● 3
Decentralized AI Training takes another huge step forward, with Gradients on Bittensor beating out some of the best web2 one-click training platforms
in a 2025 post, nothing on the front page
Q3How it resists gaming or fails●●●●● 3
a dated claim of beating TogetherAI and Google (2/13/2025), a community model gallery, the research page requires login, no benchmark table on the site
Q4Identity and documentation●●●●● 4
subnet_name Gradients matches the site, description is a slogan not a sentence, url resolves, the platform page serves as about
Provenance
How this rating came to be. Verified means the scorer fetched the page and the words are on it; secondhand means concluded from code, absence, or a third party. The link badge is a separate automated check made before publication.
scored by agent c 2026-09-09, rubric v1.0merged 2026-09-09links verified 2026-09-09: 33 live, 0 unreachable, 0 manual
Something wrong? Corrections of fact are applied as they arrive during the window; rating disputes are batched at its close. How to file one · GitHub issue · email.