โš ๏ธ Connection lost โ€” data may be stale. Retryingโ€ฆ
Live  
TAO โ€”
SN3 Alpha โ€”
Reg Burn โ€”
Emissions โ€”
Total Duels โ€”
Accepted โ€”
Queued โ€”
๐Ÿฐ Bittensor Subnet 3

The Throne Belongs
to the Best Model

King-of-the-hill pretraining. Lowest cross-entropy loss takes the crown โ€” and 100% of all SN3 emissions. No committees. No politics. Pure math.

โ›๏ธ Enter the Competition ๐Ÿ“– Read the Repo
โ€”
๐Ÿ‘‘ Current King
โ€”
Best Loss (nats/tok)
โ€”
Current Reign
โ€”
Total Duels
โ€”
King Earns / hr
Current Ruler
The King's Court

Live state of the current king and active duel evaluation.

๐Ÿ‘‘
๐Ÿ‘‘ Current King  ยท  Reign #โ€”
โ€”
โš”๏ธ Active Duel
โš”๏ธ
No active evaluation
Validator is idle or between duels
Acceptance Rate
โ€”
of all evaluations
Best Loss (King)
โ€”
nats / token
Avg Eval Time
โ€”
seconds per duel
Duels / Hour
โ€”
avg over last 6h
Progress Over Time
Loss Trajectory

Cross-entropy loss per duel โ€” king improving as challengers push the frontier down.

Avg Cross-Entropy Loss per Evaluation
King Loss
Challenger Loss
Competition Record
Duel History

All evaluations โ€” every challenger's attempt at the crown.

All ๐Ÿ‘‘ Accepted โš  Errors King Holds
Eval ID UID Challenger Verdict King Loss Chall Loss ฮผฬ‚ Time (s) When
Loadingโ€ฆ
โ€”
Pending
Evaluation Queue
โ€”

Miners waiting for their duel. Queue data is estimated โ€” API returns live evals only.

UID HuggingFace Repo Hotkey Block Queued Type
Loadingโ€ฆ
โ€”
Enter the Competition
How to Claim the Crown

Five steps to go from zero to competing for 100% of SN3 emissions.

1
Clone the repo & set up
Install Python, PyTorch, bittensor CLI, and create a HuggingFace account.
git clone https://github.com/unarbos/teutonic pip install -r requirements.txt
2
Download the current king
Read miner.py to understand how the validator fetches models. Your challenger must match the king's exact architecture.
3
Train your challenger
Any hardware, any approach. Beat the king's cross-entropy on 20,000 eval samples. Upload to HuggingFace as */Teutonic-III-* with safetensors weights.
4
Register & commit on-chain
Register on netuid 3, then set your commitment. Format: king_hash_prefix:hf_repo:model_hash (3 colon-separated parts; first 16 hex chars of king hash). Validator picks up your submission within ~30 seconds.
btcli subnet register --netuid 3 --network finney
5
Win the duel, earn the throne
Dethrone the king โ†’ 100% of all SN3 emissions flow to your hotkey every epoch, for as long as you hold the crown. Until someone beats you.
Scoring Mechanics
How Winners Are Chosen

No subjectivity. Pure math.

โš”๏ธ
Sequential Probability Ratio Test
Statistical test with ฮฑ=0.001, ฮด=0.01. ฮผฬ‚ > 0 means challenger beats king. The test runs continuously as samples are evaluated โ€” early decisive wins resolve faster.
๐Ÿ“
20,000 Evaluation Samples
Each sample is a 2,048-token sequence from the eval dataset. Avg cross-entropy loss across all samples decides the winner. Lower is better.
๐Ÿ”’
Architecture Lock
Challengers must match the king's exact architecture โ€” same hidden_size, num_layers, attention heads, vocab_size. What changes is the weights.
๐Ÿ†
100% Winner-Takes-All
All validator weights go 100% to the current king. Every epoch, for as long as you hold the crown. No partial rewards, no runner-up prizes.
The Mission
Training a Trillion Parameters,
Decentralised

Why Teutonic exists โ€” and what it's building toward.

“It took them 3 months to train a convincing 1B. We will do it in 1 week.

It took them 1 year to train a 30B. We are going to do it in 1 month.

It took them the life of the entire subnet to get to 72B. We are going to train 1 Trillion.”
โ€” Const, Teutonic architect
๐ŸŒ
Hardware-Agnostic
Teutonic is Const's evolution of Templar, which failed because it forced everyone onto the same hardware. Any compute, any training approach โ€” the only constraint is beating the loss.
๐Ÿ“ˆ
Scaling Laws in Action
Currently at 1B parameters, loss ~2.67 nats/token. Every dethronement represents a real step toward the 1T parameter frontier โ€” the largest decentralised training run in Bittensor history.
โšก
Open Competition
The throne is always open. Anyone with compute and ambition can challenge. The best model wins โ€” every time, automatically, on-chain.
Scaling Laws Roadmap
Scale Approx Loss (nats/tok) Improvement Status
~1B ~2.67 โ€” โœ… Live now
~10B ~1.60 โˆ’0.4 ๐ŸŽฏ Next target
~100B ~1.30 โˆ’0.3 ๐Ÿ”œ Roadmap
~1T ~1.10 โˆ’0.2 ๐Ÿ† The goal
Training Resources
The Eval Dataset

The validator's eval dataset is the ground truth. Train on it, understand it, beat it.

Teutonic's eval harness scores against its own curated dataset on Hippius Storage.
1,992 shards ร— ~440M tokens = ~877B tokens total (~3.7TB)
Tokenizer: unsloth/gemma-3-1b-it  ยท  Source: CulturaX
๐Ÿ’ก Strategy: The more of this dataset you train on, the lower your loss. Check the manifest for the latest shards โ€” it grows as the network runs. Training on more data = lower loss = better shot at the crown.
๐Ÿ“‹ Manifest v2 (1,992 shards) ๐Ÿ’พ Shard 0 (~440M tokens) ๐Ÿค— CulturaX Source โš™๏ธ Teutonic Repo ๐Ÿ Seed Model
โ†ป Next refresh in 30s