Teutonic is a king-of-the-hill pretraining system for Bittensor subnet 3.
SN 3. Rank 11 of 32 by emission
Teutonic
Emission2.4%of the block reward · rank 11 of 32
Participants5 miners · 3 validators
Registration0.008094 TAO
Verified2026-09-09links checked and scores merged
The programtraining · github.com@72a9f85 · read 2026-09-19
The work
A trained checkpoint of Teutonic-II, a 110 billion parameter mixture-of-experts language model, uploaded privately through a validator-issued credential and signalled ready on chain. Only the weight shards may differ from the genesis files. Each hotkey gets one submission forever. What is paid for is a reign: holding, or having recently held, the crown.
own docs · teutonic.ai
You (a miner) train a better checkpoint (a challenger), upload it through a validator-issued encrypted upload credential
How it is scored
A paired duel on the same text sequences: per sequence the challenger's loss is subtracted from the king's, a paired bootstrap of the mean difference gives a one-sided lower confidence bound, and the challenger is crowned only if that bound clears delta. chain.toml at the pinned commit sets 20,000 sequences across five datasets and delta_threshold 0.0095; copied weights are rejected before any duel.
code · raw.githubusercontent.com · teutonic/evaluation/policy.py
lcb = float(np.quantile(boot, alpha)); accepted = lcb > delta_threshold
Sources disagree llms.txt, last verified 2026-08-24, states delta 0.5 nats and n 2000; chain.toml at the 2026-09-19 commit sets delta_threshold 0.0095 and n 20000, and dashboard.json lists seven configs stepping from 0.5 to 0.0095, the current king crowned under 0.0125.
How the pay splits
Equal shares, not winner-take-all. The current king and up to four prior kings still registered each get one equal share of the weight, 20 percent at five; when none maps to the metagraph the whole weight burns to uid 0. dashboard.json on 2026-09-19 showed reigns 19 to 23 at weight 0.2 each, so one crown pays for roughly four more reigns.
code · raw.githubusercontent.com · teutonic/validator/runtime.py
share = 1.0 / len(target_uids); king_chain_size: int = 5
Who judges
Three validators were active on the 2026-09-09 chain snapshot. Commit-reveal is on with a four-epoch reveal period. Judging is one control plane: a validator service sends king and challenger to a remote GPU evaluator, records the bootstrap verdict in Postgres, and a weight publisher pushes the equal-share plan; a burn script routes everything to uid 0 when the evaluator is down. Nothing is sampled from other validators.
chain · hyperparams-2026-09-19.json
commit_reveal_weights_enabled True, commit_reveal_period 4, weights_version 2000; chain-2026-09-09.json active_validators 3
How often it turns
Emission settles every 360 blocks, about 72 minutes. Weights are republished every 101 blocks, a constraint in the control-plane schema. The puzzle turns per duel: a challenger is evaluated when it reaches the queue front and a king stands until beaten; the dashboard showed reign 23 crowned 2026-09-17. A new registration is immune from deregistration for 10,000 blocks, about 33 hours.
chain · hyperparams-2026-09-19.json
tempo 360, immunity_period 10000; dashboard.json weight_status cadence_blocks 101
Who buys
Nobody. Crowned checkpoints are published to a public bucket for free download and the August 2026 report names researchers as the audience. There is no price, no customer and no product page. The income loop llms.txt describes is emission sold for TAO to buy more GPU time and train again.
our reading · teutonic.ai
Kings earn the subnet's emissions (paid in SN3 "alpha" tokens). You sell alpha for TAO, buy more GPU time, train again.
What has been gamed
Weight copying: a challenger whose safetensors match the king is rejected at evaluation, and the same weights by SHA-256 get at most three completed evaluations across all hotkeys. Uploads sit in private prefixes and challenger identity is hidden until promotion. The evaluation sample is seeded from a block hash unknown at submission. Contract files must byte-match genesis and tensors named mtp are refused.
code · raw.githubusercontent.com · teutonic/validator/service.py
if "challenger .safetensors are identical to the king" in reason: ... return "safetensors_reuse_limit"
5 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.
What is said about it
By other voices in the map, verbatim and dated.
Bittensor's home run swing at distributed training and compute explained by it's charismatic leader himself. $TAO
It would be like investing in a startup and then the startup just says, you know, “I don't like the way capitalism works, I'm out,” and then they take the money ... yes, you should be able to leave your subnet and go do something independent if you want, of course. But, you don't get to leave with everybody's tokens
I found him a little squirrely ... I kind of pushed him on, “Hey, what are you doing next and how does this work?” And the answers, Mark, weren't as crisp or as tight as I might have expected from somebody running something significant
What it shows
Explained from zero by the index, from the subnet's own materials, for someone who has never heard of Bittensor.
SN3 Teutonic is a giant-pumpkin weigh-off for AI language models: the heaviest checkpoint stays champion until a challenger clearly outweighs it, and every champion's seeds are handed to the next growers.
The commodity, explained from zero
The pumpkin is a checkpoint: a saved copy of a language model's weights, the file needed to run it. Teutonic's current champion is a 110 billion parameter model, about 220 gigabytes, with around 7 billion active per word. Growing one means feeding it text for weeks on racks of GPUs. The subnet does not do the growing; it runs the weigh-off and pays the champion.
A checkpoint has a price because a good base model is the starting point for everything else in AI. The materials name no buyer. The August 2026 technical report says the models are for researchers studying decentralized training, and the current king's weights sit in a public bucket anyone can download, alongside an earlier 10 billion parameter model on Hugging Face. Nobody pays for them; the emission is the prize.
Why it is on Bittensor at all
The report's claim is coordination: "decentralized competition can coordinate a productive search over training data and optimization strategies without prescribing a common training pipeline". Over 70 days, 2,163 evaluations produced 203 new champions. The authors add that they cannot separate "the contribution of decentralization from data, compute, or training choices". Every champion's weights are published at once, so challengers grow from its seed. The comparison breaks on the calendar: a fair weighs everything on one day and pays once, while Teutonic weighs each challenger as it arrives and pays the champion block by block while it reigns.
How the work gets done
Miners (the growers) train a better checkpoint on their own hardware, "8x H200/B200-class or better", upload it, and commit an irreversible ready signal on chain; each hotkey gets one entry forever. Validators (the judges) send champion and challenger to a remote GPU and score both on the same text sequences, then run a paired bootstrap, a statistical test that the challenger's loss is lower by more than a threshold, at 99.9 percent confidence. If so, the challenger becomes king. Emission, the subnet's share of newly minted TAO, is split equally among the current king and up to four previous kings still registered. Copied weights get three evaluations at most, and the evaluation sample is seeded from a block hash unknown before submission.
How you would know it works
The dashboard at teutonic.ai shows the current king (crowned 2026-08-27), its benchmark scores, and every challenge in the queue; the raw data is at teutonic.ai/dashboard.json. The August 2026 report scores Teutonic-I 10B at 62.28 percent across 11 benchmarks.
What is missing
The miner documentation is one file, llms.txt, written for AI agents, and its numbers lag the live system: it states a 0.5 nat threshold and 2,000 sequences, while the dashboard data showed 0.025 on September 9 and 0.1 on September 10, with a 20,000 sequence evaluation in progress. The chain description is two words, "Coordinated Learning", and the discord field holds a handle, not a link. The contact arbos@bittensor.com appears nowhere on the site or in the repository. There is no page for stakers, no product, no price and no support path. The chain shows 5 active miners and 3 validators.
Go deeper
Sources for this explainer
Metaphor: a giant-pumpkin weigh-off for language models. 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 6/7- subnet_name
- Teutonic
- github_repo
- github.com
- subnet_contact
- arbos@bittensor.com
- subnet_url
- teutonic.ai
- discord
- @unarbos dead
- description
- Coordinated Learning
- additional
- not set
Stakers and validators●●●●● 2.8
Should I allocate here? · rank 23 of 32 for this audience
Q1What it is●●●●● 4
Teutonic is a king-of-the-hill pretraining system for Bittensor subnet 3. Miners submit immutable model checkpoints.
Output is a crowned checkpoint, no price, chain says Coordinated Learning
Q2Who it is for●●●●● 2
Kings earn the subnet emissions (paid in SN3 alpha tokens). You sell alpha for TAO, buy more GPU time, train again.
Dashboard shows reign payouts and a 16 model queue, no users or revenue
Q3How it resists gaming or fails●●●●● 3
section 9 Known exploits and defenses and section 13 failure messages, one click from the dashboard, but the file says delta 0.5 nats and n 2000 while the live manifest says 0.025 and 20000
Q4Identity and documentation●●●●● 2
identity 6 of 7, discord holds the text @unarbos not a link, contact arbos@bittensor.com nowhere on site or repo, github and url resolve, no staker page, dashboard header shows burn and alpha price
Miners●●●●● 3.8
Can I compete, and what wins? · rank 11 of 32 for this audience
Q1What it is●●●●● 4
You (a miner) train a better checkpoint (a challenger), upload it through a validator-issued encrypted upload credential, and commit an irreversible on-chain ready signal.
Full contract, for agents
Q2Who it is for●●●●● 4
For training you need a multi-GPU node (8x H200/B200-class or better) or multi-node setup
plus where to read the burn (0.007 TAO today) and a competitiveness statement, machine register
Q3How it resists gaming or fails●●●●● 3
paired bootstrap lower bound above delta in prose, one submission per hotkey forever, deregistration warning, example numbers (0.5 nats, 4.6 nat win) contradicted by the live delta of 0.025
Q4Identity and documentation●●●●● 4
README miner CLI guide matches llms.txt last verified 2026-08-24, repo pushed 2026-09-09, no tagged releases, eval parameters in the guide lag the live manifest
Buyers and enterprises●●●●● 1.5
Can I use this today? · rank 26 of 32 for this audience
Q1What it is●●●●● 3
Model Weights
link to Hugging Face and a public bucket path for the current king in llms.txt, checkpoints are the only output, no product page, no API, nothing says who they are for
Q2Who it is for●●●●● 0
no sourcesecondhandunchecked
no pricing and no customer type in any own or third party material
Q3How it resists gaming or fails●●●●● 1
dashboard panels for weight publication and service status cover validator health only, discord field is not a link, no support address, no status page for anything a buyer would use
Q4Identity and documentation●●●●● 2
no API reference, llms.txt and README are the docs, chain contact arbos@bittensor.com appears nowhere on the site or in the repo
Newcomers●●●●● 3.0
What is this and why does it matter? · rank 22 of 32 for this audience
Q1What it is●●●●● 2
Teutonic is a king-of-the-hill pretraining system for Bittensor subnet 3.
Technical opening, the site is a dashboard of numbers, the plainest account is the llms.txt TL, DR written for agents
Q2Who it is for●●●●● 3
decentralized competition can coordinate a productive search over training data and optimization strategies without prescribing a common training pipeline
vs centralized training, needs ML knowledge
Q3How it resists gaming or fails●●●●● 4
report dated August 2026 with benchmark scores, live dashboard shows ten benchmarks for the current king (MMLU 69.0 percent) and crowning dates, empty to a fetcher, numbers unexplained
Q4Identity and documentation●●●●● 3
subnet_name Teutonic matches the site, description Coordinated Learning is not a sentence, url resolves, no about or learn page (paper page only), discord field is not a link
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: 30 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.