The swarm just got faster.
We've eliminated one of decentralized training's biggest bottlenecks, speeding up merging by 5x.
We're redefining the shape of AI training.
Introducing the latest evolution of IOTA.
Live now at @ExploitSummit the Macrocosmos co-founders @macrocrux and @WSquires are on stage presenting iota now!
Distributed training across scattered hardware, coordinated as one system.
Watch live at the link below.
https://t.co/wnETRODYw6
A 16-billion parameter model is now training live on @IOTA_SN9; globally distributed over three continents, heterogeneous, permissionless, owned by no single entity.
Today, we present the next stage of Project Orion - Orion-16B.
Over the coming weeks, we'll show what it means to train across distributed, heterogeneous and "unpredictable" compute: the engineering, research and systems beneath the model, the providers powering it, and the moments that prove the infrastructure holds.
More details below.
The AI market has been actively embracing open-source models, many of which are from China. Models from this region have moved from 1.2% of global usage share in late 2024 to 30% in 2025.
Airbnb, AWS, Azure, HSBC, and GCP are just a handful of the corporations bringing these models into their architecture. All are giants in their respective industries. We have moved far beyond open-source being a hedge against centralised systems, and towards them directly shaping our future.
DeepSeek led the charge, with their R1 model being trained for $5.6 million. In comparison, Western models typically cost $80-100 million. However, it is not exclusive to them.
Qwen2.5-Max is priced at $1.60 per million input tokens and $6.40 per million output tokens on Alibaba's first-party API. Economically, this is roughly 35–45% below GPT-4o ($2.50 input / $10 output) and about half the price of Claude 3.5 Sonnet ($3 / $15).
The economics prove that models can be trained at a far lower cost, and can undercut on inference.
These benefits come directly from open-source architecture and practices, and are aligned to what @IOTA_SN9 can offer.
https://t.co/Tcnh9uv6Uo
We’re excited to unveil the latest @IOTA_SN9 results at @proofoftalk next week.
Joining @bitstarterAI on stage, we’ll reveal a major breakthrough in distributed training performance as we celebrate IOTA’s one year anniversary.
@macrocrux and @macrozack will discuss how distributed intelligence is being built, what startups need to achieve their goals in the space, and where Bittensor fits into the future of open AI infrastructure.
Date: Tuesday 2 June
Time: 13:00–14:00 CEST
Location: Proof of Talk
Finally we released an overview demonstrating the last 12 months of our work on distributed training. The last year was about making the system reliable & fast, the next year will be all about scaling up the models.
Training frontier models over the internet requires new techniques.
Today, we present ResBM, a residual encoder-decoder bottleneck architecture that enables 128x activation compression for low-bandwidth distributed pipeline parallel training.
Developed for @IOTA_SN9, we show SOTA compression without significant loss in convergence rates, increases in memory, or compute overhead.
Expect the full paper release in the next 72 hours.
Guess what... more spending on data centres, based on enormous borrowing. When the capital constraints hit, heterogenous systems like @IOTA_SN9 will be the only solution
Tomorrow, Training at Home goes live 🚨 Livestream on X @ 8pm, in-person event in London 🇬🇧
@IOTA_SN9 is opening its network to the world, allowing you to train our model with just a MacBook
🕖 7pm-10pm GMT
📅 Tue December 9th
📍 MindSpace, 142-146 Old Street + streamed on X
Bittensor Guru Podcast S2E13 Subnet 9 @IOTA_SN9 by @MacrocosmosAI - Incentivized Orchestrated Training Architecture allows anyone to mine on Bittensor and contribute compute to its global AI training cluster. $TAO
🚨 JUST IN: Anyone with a Macbook will be able to mine on IOTA (SN9) @IOTA_SN9.
@KeithSingery - "I predict IOTA mining will onboard more users into Bittensor than all current sources combined."
Podcast lands on Monday 🔥
That's a lot of borrowing... imagine what might be saved if they could access and effectively deploy the long tail of "obsolete" and underutilised compute reachable through @IOTA_SN9 for the same goals
Join us to train AI models over the open internet with nothing more than a simple app download. Become a Bittensor miner with no coding, no btcli, no complex set up.
Launching soon from @MacrocosmosAI and @IOTA_SN9
Today, we are announcing the launch of Training at Home, our effort to democratise the creation of AI models to allow the many to participate, not just the few.
Training at Home is an easy to use application that allows anyone with a Macbook to contribute to creating AI models. No complex coding, no complicated set-up - just install and train.
This is the next step of iota’s mission to scale distributed training until it is indistinguishable from training in a data centre.
Together, we can train models collectively, and you can help us shape a more open, democratic, and accessible future for AI.
Contribute, compute, own the future.
Coming soon.
Frontier AI capex spend is increasingly absurd - 10GW data centre campuses need the power output of a small country. IOTA offers a different solution:
- Access to underutilised resource = greater flexibility, lower cost
- Hardware agnostic = open incentivised scaling, lower barrier to entry
- Geographically distributed = more energy efficient
@IOTA_SN9 is on the path to making AI model training economically viable for the long term and capturing a piece of this multi-billion dollar market
Training AI models is hard, costly, and difficult to decentralize.
The biggest tech companies are pouring billions of dollars into vast datacentres that might be out of date by the time they’re built.
Big Tech’s vast datacentres cost billions - but could be out of date by the time they’re built.
The world needs an alternative.
IOTA allows anyone, anywhere to train models collectively - combining compute from all over the globe, large or small, to create a competitive advantage over Big Tech.
So that everyone can participate in pretraining the latest models, paving the way for a democratized, decentralized future, built on Bittensor.
And soon, collectively solving one of AI’s most complex problems will be as simple as clicking ‘Contribute Compute.’
Designed by @MacrocosmosAI - trained by anyone, anywhere, anytime.
⚡️ Inventive Mechanisms - live right now ⚡️
Featuring our first guests: @tsliceAI x KaiserMerkle
Wanna learn how you power agents with SN13? These two know all about it 🧩
https://t.co/H0i9sxrs3M
Will's working on what I believe to be one if the biggest opportunities in AI: distributed model training
IOTA (Subnet 9) is the only *live* distributed training network
"If we project IOTA forward, we will have a system that we could actually train a frontier model on, that supplies compute for the world’s most valuable task at a fraction of the cost"
@KeithSingery has been a friend and supporter since the beginning. The Embryon plan is testament to his unwavering belief in #Bittensor - there's no one better to see #SN25 through its next phase
I have acquired Bittensor Subnet 25 from @MacrocosmosAI. It is now called Embryon, the embryonic incubator for the next great team on Bittensor.
You're used to being a $TAO holder chasing subnets; Embryon turns you into an alpha holder with the subnet teams chasing you. 🥕
Investing in subnets is now as easy as shopping on Amazon.
Design your own strategy now and let Mentat optimize it for you.
Discover Mentat+ : https://t.co/IMD9cQtBU7