Teutonic

Teutonic is a decentralized pretraining platform operating as Subnet 3 (SN3) on the Bittensor network. The platform runs a king-of-the-hill competition where miners train large language model checkpoints and compete to produce the single best model on the subnet. At any moment one checkpoint holds the title of king, and miners earn rewards by submitting a stronger challenger that beats it on a held-out evaluation. The project is operated under the Arbos organization and is published openly at teutonic.ai.

The core mechanic is a direct head-to-head duel. A miner uploads a trained model to Hippius Hub, a decentralized storage layer, and commits an immutable fingerprint of that exact checkpoint on-chain. A validator then pits the challenger against the reigning king by measuring how well each model predicts the next word across thousands of unseen text sequences. If the challenger is statistically better by a clear margin, it becomes the new king and starts earning rewards.

Teutonic runs on open-source large language model architectures in the Qwen3 family. Every challenger is pinned to a fixed model shape and a verified snapshot, so miners compete purely on training quality rather than on changing the underlying model, which keeps the contest measurable and resistant to manipulation.

A miner starts from the current king checkpoint, then trains a challenger model and uploads it to Hippius Hub, which returns an immutable digest in the form sha256:<hash>. The miner submits an on-chain reveal as v4|repo|digest|author_hotkey. That digest is the binding commitment, so the validator evaluates exactly the checkpoint the miner locked in and nothing swapped in later.

Before evaluation, the validator runs integrity checks. The repository must follow the required naming pattern and include the first eight characters of the miner's coldkey to prevent impersonation. The model configuration must match the king on every structural property including vocabulary size, hidden size, layers, and attention heads, ship clean safetensors files with no custom code, and stay under the operator size cap.

The evaluation is a paired cross-entropy duel on a dedicated GPU server. Both models are scored on the same token sequences, selected through a seed derived from the reveal block hash and miner hotkey so the test set cannot be gamed. The system measures next-token loss across roughly 5,000 sequences, then runs a bootstrap to confirm the result is real. A challenger is crowned only if its lower confidence bound clears 0.0025 nats per token.

Rewards are not winner-take-all. The validator keeps a rolling chain of kings, splitting emissions equally across the current king plus up to four previous distinct kings that remain registered, so two kings split 50/50 and five split 20 percent each.