Hux

Building #bittensor subnets 3 and 66 with @DendriteHQ

Pushing further: the unproven:

[SOMA @SomaSubnet]

SOMA × GM is live.

We compress context and route agent requests to @say_gm_. They provide the model inference.

Two Bittensor subnets working together, with real traffic flowing through the integration.

Compress. Pay less.

[Hux @hux_dendrite]

See you at the conference in Canada.

I'll be leading the technical part of the Mining 101 workshop on the second day.

I'm going to introduce new folks to mining on Bittensor and share our perspective on mining in a short and precise way.

We'll also cover real world examples and do some hands on practice.

[Exploit Summit @ExploitSummit]

. @hux_dendrite works across LLM training, agents + production AI.

He’s mined across 15 #Bittensor subnets and co-authored Teutonic-I 10B, built through SN3’s 70-day decentralized training competition with 2,163 model duels and 203 successful improvements.

Now he’s leading our Bittensor Mining 101 session at Exploit:
https://t.co/EbNmi3nRf6

[Hux @hux_dendrite]

@subnet_mania Albedo is used to evaluate and verify whether knowledge is moving in the right direction

[Hux @hux_dendrite]

Knowledge and comprehension are what drive the further progress of https://t.co/KliUyziPoI

[const @const_reborn]

we will need compute (@lium_io, @TargonCompute)
we will need inference (@engyai, @chutes_ai)
we will need data (@Hippius_cloud)
we will need knowledge (@teutonic)
If we are going to have free intelligence (@affine_io)

[Hux @hux_dendrite]

@const_reborn @lium_io @TargonCompute @engyai @chutes_ai @Hippius_cloud @teutonic @affine_io That's a spot on

[Tony @tony_dendrite]

Speed is key, so I just made our evals 10x faster.

Added Flash Attention 4 support, batched the samples, and fixed a bug in Transformers along the way. Teutonic evals went from 1h down to 20min including model downloads. Big win for iteration speed.

[Conjectures.io @conjectures_io]

https://t.co/w6tub1chxY

[SOMA @SomaSubnet]

SOMA is live for GitHub Copilot.

All Early Access emails have now been sent. If you signed up, check your inbox.

You can plug SOMA into Copilot and use DeepSeek V4 Pro with 10% savings. Every Early Access user gets $5 in credits to test it on real coding tasks.

There are no SOMA platform fees during Early Access.

If you missed Early Access, we’ll open SOMA to everyone soon.

[Hux @hux_dendrite]

1 WEEK IN. 186 MODELS EVALUATED

The benchmarks are climbing, new models keep entering, and we’re just getting started.

In the last week, we added new datasets to the evaluation suite. Now we’re evaluating on more than 7T tokens. And this is only the beginning.

No invite needed. Just bring your model and see how far you can push it.

More info about competition you can find: https://t.co/AhLCS6n6dD

[SOMA @SomaSubnet]

The best product feedback is a production deployment you didn’t ask for.

@say_gm_ ported SOMA’s OpenClaw compression core to Rust and is running it inside their privacy-first gateway for dozens of AI models.

That’s what we built SOMA for.

[Tony @tony_dendrite]

110B changes the scale, not the rules. How it works:

The validator encrypts a prefix-scoped, TTL-bounded R2 token to your Ed25519 hotkey and leaves it in a public mailbox. Anyone can fetch the ciphertext, only you can open it, and it writes to your prefix and nowhere else. Then `ready` commits a signed manifest on chain. Irreversible, one shot per hotkey.

Eval is paired. King and challenger score identical sequences, drawn by blake2b(finalized block hash, your hotkey), so you don't pick your own data. Per-sequence cross-entropy on finewebedu, resampled to get a conservative lower bound on the mean improvement. You take the crown only if that bound clears delta, which is 0.5 at launch since early gains off a fresh init are large. Ties go to the king.

Architecture is locked to the king's config. The only lever left is training a checkpoint with lower loss.

https://t.co/1rIR97LYyu

Nearly a full rewrite since the last competition. The mechanism survived it intact, which says the design was right. Everything underneath it had to change to move checkpoints this size. @DendriteHQ

[Hux @hux_dendrite]

@const_reborn The journey has just begun, and we're waiting for talented miners to join and push this competition to its limits.

[Hux @hux_dendrite]

Here we go!

The 110B permissionless training competition has just gone live on Teutonic.

We've spent a lot of time exploring different architectures and configurations, and ultimately decided to base our model on the MiMo-V2 architecture.

We can't wait to see the loss going down!

[const @const_reborn]

Announcing Teutonic-II-110B: a frontier permissionless training of a 110B LLM

We are scaling 10x beyond our previous model and 50% larger than any prior decentralized run.

[Dendrite @DendriteHQ]

70 days. 2,163 duels. 203 coronations. One 10B model left standing.

Teutonic-I beat Covenant-72B's 11-benchmark average with 7x fewer parameters, without synced gradients or a fixed recipe.

@ch4ris4rd on the system he built with @hux_dendrite, @tony_dendrite & @const_reborn.

[Dendrite @DendriteHQ]

The new king of decentralized AI has just been crowned.

A 10B model just beat 72B.

Teutonic-I 10B: 62.28%
Covenant-72B: 57.55%

Same 11-benchmark average.

SN3 has changed hands. @const_reborn is the owner.

This is only the beginning. A new chapter starts here.

🧵 1/8

[Hux @hux_dendrite]

So proud to be part of this team with @DendriteHQ. Transparency, fairness, and permissionless participation were key to achieving these results.

We’ve achieved a lot, but the best is yet to come.

[const @const_reborn]

Today, after 4 months, Subnet 3 (previously @tplr_ai) got retaken by holders through vote.

We're celebrating by releasing our first model.

Teutonic-I: the worlds most performant decentralized LLM.

https://t.co/TNHtP0hfyr

[Dendrite @DendriteHQ]

On Albedo/SN97, we’ve moved most of its LLM traffic from OpenRouter to @engyai - 100% of DeepSeek and 40% of GLM

Worst-case request stalls dropped from 75s to under 2s. Evaluations are now 10-30% faster

Great work by @AlgodTrading, @totheagi and the Engy team

[Pioτr_Barbachowski @pbarbachowski]

For my first post, I'm quoting a letter on why open models matter, because we've been betting on exactly that since 2022.

Back in 2022, five of us spun up our first virtual machine. No IT background between us, and the first open-source model we launched broke within an hour.

Today we're @DendriteHQ, a 50-engineer team building on Bittensor: we own one subnet (@SomaSubnet), help build three others, and run mining operations on five.

All of it on open models, most of it through a bear market.

Which honestly turned out to be the useful part.
When there's no hype around, you find out fast whether your work is real.

The letter says the world needs frontier open models.
It also needs places where open models prove themselves in production, daily, with money on the line. That's the layer we work in.

[Jensen Huang @JensenHuang]

Attackers have frontier AI. Defenders need a frontier AI ecosystem—the best open and closed models, force-multiplied by a global community.

During the Hugging Face incident, closed AI blocked essential forensics. An open-weight frontier model helped contain the intrusion.

That’s why we created the Open Secure AI Alliance.

[Hux @hux_dendrite]

@nvidia @JensenHuang Have you heard of #Bittensor?