Teutonic Is Live. Const Is Cooking. The 1 Trillion Parameter Run Is Loading.

Four days after Covenant AI walked away from Subnet 3 and dumped $10 million on the community, Const didn’t just revive the subnet. He rebuilt it from scratch.

Meet Teutonic.

SN3 has a new name, a new mechanism, and a new mission. And if Const’s Discord messages are anything to go by, the ambition has gone up by several orders of magnitude.

“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.” – Const

What Is Teutonic?

Teutonic is Const’s complete rebuild of decentralized LLM pre-training on Bittensor. He didn’t patch Templar. He replaced it with something more powerful and ambitious.

The core concept is king-of-the-hill. Here’s how it works:

Validators evaluate challengers against 10,000 samples. The model with the lowest cross-entropy loss wins the crown. The reigning king captures 100% of SN3 emissions every epoch. You want to earn? Beat the king’s loss on its exact architecture. Dethrone it.

The seed king launched on April 13, 2026, a 0.9B parameter Gemma3 model in BF16. And it’s already been through 11 reign changes. The Teutonic-I dashboard shows the loss dropping from ~13 down to the low 5s across 372 evaluations. The perplexity chart is trending in one direction: down.

Why Templar’s Mechanism Had to Go

Const didn’t just build something new, he publicly explained why the old mechanism was broken. And he should know since he wrote the original Templar mechanism himself.

“I think their mechanism — which I wrote originally, by the way — is limited and won’t work,” he said on Discord. “Templar was super limited because it forced everyone onto the same hardware and they were very limited in where they could optimize.”

Templar required miners to use identical hardware configurations. Everyone was running the same setup, which meant the only variable was compute time and not creativity, not architectural innovation, not data quality. The ceiling was low.

Teutonic blows that ceiling off. The new design is hardware-agnostic. Any GPU, any training approach, any compute stack. The evaluation doesn’t care what hardware you used or what training method you applied. It only cares about one number: your cross-entropy loss on the eval set.

“Effectively we structure it so that the next update to the model is a competition,” Const explained. “This allows scaling and optimization around that eval.”

The result is a system that rewards clever tricks (like better data curation, architectural innovations, training efficiencies), not just brute force. At scale, that distinction matters enormously. The jump from 1B to 1T parameters won’t come from throwing more of the same hardware at the problem. It’ll come from someone finding a breakthrough that the king-of-the-hill mechanism is specifically designed to incentivize.

Where It Stands Right Now

As of today, Teutonic is training a 1B parameter model. The dashboard at projectnobi.ai/teutonic3 tracks every reign change, every duel, every step down in loss.

The current king’s perplexity sits at 5.7477 after 372 evaluations. Challengers are queuing up. The loss curve is heading down. The mechanism is working.

But Const is being clear-eyed about where things stand. He’s told the community this is a testing phase at 1B. The proper evaluation design with bounding box updates is what will allow scaling to 1T. The 1B run is proving that the mechanism works. The bigger runs come next.

The scaling roadmap maps out the path:

• ~1B parameters → ~2.0 nats/token loss (where Teutonic is now)
• ~10B → ~1.6 loss — next target
• ~100B → ~1.3 loss — on the roadmap
• ~1T → ~1.1 loss — Const’s stated goal

Each 10x in compute buys diminishing returns in loss reduction. That last stretch from 1.1 to 0.8 nats/token may require architectural breakthroughs, not just brute force. And that’s exactly what Teutonic’s competitive mechanism is designed to surface.

Important: Don’t Buy SN3 Yet

Const has been consistent and direct about this. The owner key for SN3 is still with Sam Dare. Until that’s resolved, any buys into the subnet’s alpha token could effectively put money into a pool that Dare can still drain.

“People please do not buy this subnet while it is in this state, literally just do nothing while we sort this out.” Const said on Discord. “If you are a holder and want to wait it out, thank you.”

The mechanism is live. The training is running. The results are coming in. But the token situation needs to be resolved before SN3’s alpha is a safe investment again. We will provide updates as the situation advances.

What This Means for Bittensor

Step back from the technical details, and the timeline Const is laying out is extraordinary.

Covenant took the entire life of the subnet to train a 72B parameter model. It was a historic achievement, endorsed by Jensen Huang (CEO of Nvidia), cited by Jack Clark (co-founder of Anthropic), and responsible for a 90% rally in TAO token. Nobody is taking that away.

But Teutonic is designed to move faster, scale harder, and do it through open competition rather than a single team’s coordination. A 1B model in the first week. A 30B in a month. And a 1T parameter model as the long-term target, trained permissionlessly across a decentralized network, with every improvement earned through king-of-the-hill competition.

If that works, it doesn’t just vindicate Bittensor after the Covenant exit. It makes the Covenant exit look like the event that forced the network to become what it was always supposed to be.

Do you have what it takes to train 1 trillion parameters? Join the mining race:

Official: teutonic.arbos.life · GitHub: github.com/unarbos/teutonic · Seed model: huggingface.co/unconst/Teutonic-I

This article is for informational purposes only and does not constitute financial advice.