Miner leaderboard
Who is online, and whose models customers kept. Outside miners can enter the base tier now; customer jobs open once miner terms are live.
- Online now
- 3 / 5
- Jobs
- 96
- Models selected
- 11
- Current base
- d3-mini 4.5B
Three ways to earn
Testnet allocation This validator allocates 50% to base improvement, 20% to features and 30% to customer adapters. Mini/lite split 80/20 within base and box budgets: 40/10/16/4/30 across mini base, lite base, mini boxes, lite boxes and customer adapters. Empty shares burn.
Base tier
50%Open on testnet
Improve the whole model; the winner becomes the base.
- To beat
- Your tier’s base
- Pays
- Last five kings
- Entry
- Full model weights
Feature tracks
20%Open on testnet
Teach the base a new skill. Each skill is its own competition.
- BoxesSynthetic-only pilot · mini / lite · no kings yet
- VideoLater · not designed yet
- Entry
- An adapter on the exact parent
- Pays
- Last five distinct kings per track
Customer adapters
30%Admission off for now
Train the models customers keep.
- Scored on
- Each job’s held-out test
- Opens to all
- Once miner terms are live
Graduated into the base None yet. A synthetic pilot crown is separate from graduation, which requires real-image qualification and the full base regression review.
Live from the Zils training API · refreshes every 30 secondsUpdated 05:47 UTC · sort by any column
How to read the leaderboard
Selected is how many of the miner’s candidates a customer job kept as its model, out of the jobs it was given. Jobs are customer training jobs assigned to the miner; evaluated candidates were scored on the job’s held-out test set.
Skill compares a candidate’s probability error (Brier) with guessing evenly between the answers: 0 is no better than guessing, 1 is perfect, so it compares across different jobs.
Online checked in within the last 45 seconds. Seen recently checked in within the last hour. Offline has not checked in for longer than that. Never seen registered but never checked in. Disabled gets no new jobs. Miners that are not online are dimmed.
Models come from each miner’s latest signed check-in, including while it is busy. Inactive rows show the last reported models. These are training-worker UIDs, not on-chain UIDs; a listed model does not imply customer serving is available.
Subnet 579 runs on the Bittensor testnet. Outside miners can already compete in the base tier there (below). Customer jobs stay with Zils miners until miner terms and intake are ready.
Beat the base.
Every tier has a base to beat. Win a sealed exam, then win again on a fresh one.
Start mining
curl -fsSL https://zils.ai/install.sh | shInstalls the zils and btcli commands, then zils setup asks for your wallet and Hugging Face repo. After that, zils doctor checks each step and prints the command for the next one: register on testnet 579, train a candidate, commit it. Training needs Linux x86_64 with an NVIDIA GPU. Rerun the command to update.
What we have tried
Current base
d3-mini 4.5B
- Profile
d3-mini-4.5b-v1- Input limit
- 4,096 tokens
- Inputs
- Text, up to 4 images, video
- Channels
- stable d3-mini 4.5B, latest Zils Base r5 (d3-mini 4.5B)
Live from the Zils training API.
How a new base wins
- Submit. A miner uploads replacement weights for the base to Hugging Face, then commits them on chain with
zils commit-candidate. - Score. Validators score it on a sealed exam the miner never sees, against the current base.
- Confirm. It must win again on a second, fresh exam, and customer-style adapters trained on it must not get worse.
The full rules are on the research page and in the miner guide.
What we have tried
15–16 Oct 2026
Round zero: can full-weight training beat adapters?
Not yetWe retrained every weight of the old base, H2O Lightning 4B, on one rented B300. It worked, but at equal data it trailed the small adapters we already use. No new base qualified, and the weights were not kept.
Test accuracy on ABCD, 4,096 training examples - Full weights
- 79.8%
- Adapter (LoRA)
- 84.4%
16 Oct 2026
A new candidate on the sealed exam
Cleared the first drawStock d3-mini, an open 4.5B model on the same backbone, sat the sealed round-one exam: 1,166 test cases in six families, scored by the validators’ own code. It beat the incumbent and was marked eligible.
Brier score, lower is better - H2O (incumbent)
- 0.184
- d3-mini
- 0.142
16 Oct 2026
Do adapters on the new base hold up?
Matched or betterTwo training seeds on d3-mini matched the H2O adapter on text (83.4% and 83.6% against 83.0%). On product-inspection photos, a d3-mini adapter scored 169 of 200 against 168 for the image model it would replace.
Text adapter, test accuracy on ABCD - H2O adapter
- 83.0%
- d3-mini adapter
- 83.6%
17 Oct 2026
The confirmation exam
A tie, so no winnerThe rules ask a winner to repeat its result on a second, fresh sealed exam at least 12 hours later. It did not: the two bases tied, well short of the 0.01 improvement the rules require. The adapter check passed (Brier 0.2515 against 0.2801).
Brier score on the second exam, lower is better - H2O (incumbent)
- 0.1794
- d3-mini
- 0.1790
17 Oct 2026
d3-mini becomes the base, by decision
AdoptedWe adopted d3-mini anyway, for what it adds: adapters that train better, up to four images and a video per call, and yes/no, choice and score answers. This is recorded as an override of the promotion rule, not a win under it. H2O and the separate image model stopped serving the same day.
Next
Round two: beat the base
Open on testnetChallenge the current base in your tier. Anyone registered on testnet 579 can enter: upload to Hugging Face, commit on chain, and every validator pulls and scores it. We also want the exam to cover images and video, since that is what the base is used for.
One rented card per run, teacher-authored exams rather than customer data, and one reference job in the adapter check where the rules ask for three or four. Results are recorded as measured, with their dates.
Testnet rehearsal · 25 Sept 2026
Bittensor · Recorded testnet research
Bittensor mining.
How miners train decision models, validators evaluate them, and weights reach the chain. One testnet rehearsal, from 25 September 2026, is recorded here with its evidence and limitations, followed by the base tier, where miners compete to replace the base model itself.
Training & evaluation
Train. Submit. Evaluate. Reward.
- 01
Train
Miners train compatible adapters and decision heads with a fixed one-epoch recipe and different seeds.
- 02
Submit
Each miner freezes a candidate and signs its checkpoint hash for the validator.
- 03
Evaluate
The validator verifies the checkpoint and runs its own evaluation of the model’s probabilities.
- 04
Reward
Family-macro Brier skill determines proposed weights across eligible candidates.
One testnet round completed
Training, signed submissions, evaluation, and revealed chain weights are recorded below. A live submission queue and round feed are not connected.
View the verified roundThe local loop works on Apple Silicon and an RTX 4090. Accuracy and latency are diagnostics, not separate reward components. The subnet’s family-macro Brier and JevBench Brier use different aggregation rules.
Evaluation contractParticipants
Three miners. One validator. One host.
Three registered miners and one validator completed the recorded subnet 579 rehearsal.
- Miners · UIDs 1–3
- Trained and submitted three distinct checkpoints
- RECORDED PARTICIPATION
- Validator · UID 0
- Evaluated the checkpoints and published weights
- RECORDED PARTICIPATION
One operator ran all four processes on one host. Services exited after the round; these counts do not indicate current availability or independent operators.
Miner guideTestnet publication
Bittensor testnet · Subnet 579
First training-to-chain round completed. Revealed weights verified.
Round evidence (JSON) ↗- Network
- Bittensor testnet
- Publication receipt
- 8081301-0008
- Verified at block
- 8,081,345
- Round status
- Completed · recorded rehearsal
| Miner UID | Correct / 224 | Requested weight | On-chain value |
|---|---|---|---|
| 1 | 183 / 224 | 34.09% | 65,535 |
| 2 | 180 / 224 | 33.60% | 64,598 |
| 3 | 180 / 224 | 32.31% | 62,125 |
Weights are based on family-macro Brier skill. The chain stores maximum-scaled integers; their normalized proportions matched the requested allocation within quantization tolerance. Verified . This is the observation time, not a live refresh.
Each 0.8B checkpoint trained for one epoch on 224 examples, calibrated on 112 questions, and was evaluated on the same 224 test questions. This reused synthetic development benchmark is separate from the public JevBench comparison. All compute ran on one Apple M4 Pro with 24 GiB memory, with GPU jobs serialized.
This closed rehearsal demonstrates the training-to-chain path. It does not establish mainnet deployment, miner earnings, open competition, or automatic model promotion.
Base tier · status as of 19 October 2026
A competition for the base model itself.
Customer adapters train on one pinned base model. Since 17 October 2026 that base is d3-mini 4.5B. In the base tier, miners submit a full-weight replacement for that base. Validators score each candidate on a sealed holdout against the incumbent, and a candidate is promoted only after it also passes a second sealed draw and an adapter regression suite.
- King of the hill
- Only kings are paid. A winner becomes king when the owner promotes it. The base-tier share splits equally over the last five distinct kings still registered.
- No king
- An empty track’s share is burned.
Today’s base is stock d3-mini, a vendor base with no miner behind it, so there is no king yet. No candidate has been promoted, and nothing runs on mainnet yet.
Full promotion rules