IOTA is a framework for pretraining large language models across a network of heterogeneous, unreliable, permissionless and token incentivized machines.
SN 9. Rank 13 of 32 by emission
iota
Emission2.3%of the block reward · rank 13 of 32
Participants3 miners · 10 validators
Registration0.0005 TAO
Verified2026-09-09links checked and scores merged
The programtraining · github.com@22d702a · read 2026-09-19
The work
Forward-pass work on one slice of a language model. An orchestrator run by Macrocosmos assigns each miner one layer of a model being pretrained; the miner computes activations, the numbers passed between layers, streams them through a shared bucket, and periodically uploads its weights. What is paid for is validated activation work, not a finished model.
own docs · raw.githubusercontent.com
Miners compete to process as many activations as possible in the training stage.
How it is scored
Validators score each miner's forward-pass activations in short windows on several checks: consistency over time, agreement with peers on the same layer, expected signal shape, and anti-gaming patterns. Scores sum over a rolling window of about a day. On-chain weight is the miner's share of its run's total score, times one minus the run's burn factor, times the run's allocation.
own docs · docs.macrocosmos.ai
validators independently measure the forward-pass activations each miner produces during training and assign a score for the work done in each validation window
Sources disagree The docs say the exact thresholds live in the code; the repository at the pinned commit holds no scoring function. task_execution.py loads validation functions from an orchestrator directory that is not in the repository, and weight_setting.py fetches finished scores from /validator/global_miner_scores.
How the pay splits
Proportional. Every miner with positive verified work gets its share of its run's score; the validator normalizes the vector to sum to one. Whatever miners do not earn is set as weight on the owner UID 209, the burn. The orchestrator decides the burn share; settings.py defines FALLBACK_BURN_FACTOR = 0.8 but nothing in the repository reads it.
code · raw.githubusercontent.com · src/validator/src/validator/utils/weight_setting.py
raw_weights = torch.nn.functional.normalize(scores, p=1, dim=0); burn_factor = next((weight for uid, weight in weights.items() if uid == common_settings.OWNER_UID), None)
Who judges
Ten validators were active on the 2026-09-09 chain snapshot. Commit-reveal is off; weights may be set once per 100 blocks. A validator fetches validation tasks from the orchestrator every minute and runs them, but the scores it submits are the orchestrator's global miner scores, fetched whole; if that call fails it copies the stake-weighted weights already on chain.
chain · hyperparams-2026-09-19.json
commit_reveal_weights_enabled False, weights_rate_limit 100, weights_version 4062; chain-2026-09-09.json active_validators 10
How often it turns
Emission settles every 360 blocks, about 72 minutes. The validator submits weights every 21 minutes, on a timer independent of training epochs. Scores aggregate over a rolling window of about a day; kick policies apply at the end of each training epoch. A new registration is immune from deregistration for 5,000 blocks, about 16 hours.
chain · hyperparams-2026-09-19.json
tempo 360, immunity_period 5000; validator settings.py WEIGHT_SUBMIT_INTERVAL: int = 60 * 21
Who buys
Nobody is named. The front page on 2026-09-19 sells two routes: an SDK to run your own training on IOTA, and a managed route where you bring a workload and talk to the founding team. There is no price, no customer and no checkpoint for download. The index's dashboard fetch on 2026-09-09 read zero on every counter.
own docs · iota.macrocosmos.ai
Bring a defined workload, and we take it from qualification through to a result you can act on.
What has been gamed
The docs name the prior design's failure: under winner-takes-all, miners hoarded whole models, which IOTA answered by splitting the model into layers and paying per validated slice. Against faked work, scoring includes checks for tell-tale repetition or shortcuts; miners can be kicked each epoch by a bottom-by-score or threshold policy, removed by an operator, or banned by tier.
own docs · docs.macrocosmos.ai
every miner had to fit an entire model locally, and "winner-takes-all" rewards encouraged model hoarding
3 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. All 15 quotes in the map.
Their subnet
Subnet 9 is the heart of Bittensor's vision, then the home of the first trustless swarm pretraining protocol where participants are rewarded as a collective.
The pre-training subnet (SN9), is at the heart of Bittensor's vision. Pre-training is the crucial first step in the creation of all modern AI models, where enormous amounts of data and computational power are consumed by the model in order for it to begin to understand the world.
Today on SN9, the pre-training process is distributed between highly skilled teams which compete to continuously improve models and get paid to do so. As a result SN9 produces world class large language models (LLMs) that are truly open: open source, open data, open models, open competition.
Subnet 9 proved decentralized LLM pretraining is viable. We are proud to release the technical primer for IOTA in advance of mainnet launch on June 2nd.
Subnet 9 is now the home to the world's first ever model parallel and data parallel incentivized trustless pretraining protocol
Swarm refers to multiple nodes working towards the same goal - but IOTA harnesses Bittensor to overcome the challenges of trustless incentivization and reward
Bittensor, in their words
Macrocosmos speaks in the foundation's founding-mission register, calling Bittensor an incentivisation network, a global incentives machine, and a network of commodities whose edge is alignment with alpha holders.
The pre-training subnet (SN9), is at the heart of Bittensor's vision. Pre-training is the crucial first step in the creation of all modern AI models, where enormous amounts of data and computational power are consumed by the model in order for it to begin to understand the world.
There is this sort of constellation effect, of you needing lots and lots of pieces of the AI process to be really effective in this space, and in Bittensor. Ultimately, the network needs lots of these commodities to be super valuable. And so we saw a lot of harmony between building more than one of them
Swarm refers to multiple nodes working towards the same goal - but IOTA harnesses Bittensor to overcome the challenges of trustless incentivization and reward
IOTA is also a significant step towards Bittensor's original intention of creating a competitive, decentralized alternative to proprietary models
Frames collective swarm constellation heart home machine north star
What is said about it
By other voices in the map, verbatim and dated.
Anyone can now mine on Bittensor using hardware you already own by contributing compute to IOTA's global training cluster, the world's first permissionless pipeline parallel training architecture.
What it shows
Explained from zero by the index, from the subnet's own materials, for someone who has never heard of Bittensor.
SN9 IOTA builds one large language model on an assembly line whose stations sit in strangers' garages, paying each station for the parts it passes down the line rather than for a finished car.
The commodity, explained from zero
The commodity is training: the process of teaching a language model by pushing text through it and adjusting billions of internal numbers until its guesses improve. The finished product is a checkpoint, a file of those numbers that anyone could load and run. The README calls IOTA "a framework for pretraining large language models across a network of heterogeneous, unreliable, permissionless and token incentivized machines." The current model is a 1.5 billion parameter design, cut into three sections.
IOTA's bet, in the front page's words, is "to show that distributed, heterogenous, permissionless training of models can be a competitive economic alternative to training in a centralised context." The price would be whatever a lab saves by not owning the data center.
Nobody buys the output today. The materials name no customer, no price, and no checkpoint for download. The only offer on the site is on the supply side: a Train at Home app that lets a person with a Linux machine and a GPU plug in and earn. Here the comparison breaks: an assembly line exists to ship cars, and this line has not yet said who will drive the one it is building.
Why it is on Bittensor at all
The claim is that a training run can be paid for continuously, by measured contribution, rather than from one lab's payroll. The July 2025 paper says the earlier design of this subnet had each miner train a whole model alone and "rewards favored hoarding," and that IOTA splits the work so a contributor needs only a slice. Whether that beats a data center on cost is what the project says it is trying to show; no cost figure is given.
How the work gets done
An orchestrator, a coordinating server run by the team, hands each miner one layer of the model and streams activations, the intermediate numbers passed between layers, through a shared storage bucket. Miners "compete to process as many activations as possible," and periodically upload their weights. Validators measure each miner's output for consistency over time, agreement with peers on the same layer, the expected shape of real training, and "tell-tale patterns of repetition or shortcuts that suggest the miner is faking work." A miner's reward is its share of the run's total score, times the run's allocation, times one minus a burn factor, over a rolling window of about a day; the lowest scorers on a layer can be removed each epoch.
How you would know it works
The dashboard at https://iota.macrocosmos.ai/dashboard is live, but every counter read zero on 2026-09-09 while the front page said "Project Orion is Live!" The technical paper at https://arxiv.org/abs/2507.17766, dated July 2025, is the only published result.
What is missing
The README's two documentation links return Page Not Found, so there is no working miner guide. The dashboard reads zero. The chain description, "Bringing liquid training to the world," appears nowhere on the site. Support is Discord only, and the last tagged release is v3.0.0 from 2026-03-09.
Go deeper
- Front door with stated objective: https://iota.macrocosmos.ai/
- Scoring, rewards and kicking page: https://docs.macrocosmos.ai/subnets/subnet-9-iota/scoring-rewards-and-kicking.md
- The technical paper, July 2025: https://arxiv.org/abs/2507.17766
Sources for this explainer
Metaphor: an assembly line whose stations sit in strangers' garages. 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
- iota
- github_repo
- github.com
- subnet_contact
- hello@macrocosmos.ai
- subnet_url
- iota.macrocosmos.ai
- discord
- discord.gg dead
- description
- Bringing liquid training to the world
- additional
- not set
Stakers and validators●●●●● 3.0
Should I allocate here? · rank 18 of 32 for this audience
Q1What it is●●●●● 3
distributed, heterogenous, permissionless training of models can be a competitive economic alternative to training in a centralised context
output is a pretrained LLM, no unit or price
Q2Who it is for●●●●● 2
Loss 0.0000 Tokens 0.00B Active nodes 0 Parameters 0.0B
on the live dashboard while the front page says Project Orion is Live, no customers or revenue
Q3How it resists gaming or fails●●●●● 4
named Scoring, Rewards and Kicking page two clicks from the docs root with consistency, peer agreement and anti-gaming checks, a weight formula and a kick policy, thresholds left to the code
Q4Identity and documentation●●●●● 3
identity 6 of 7, github and url resolve, contact matches the site footer, chain discord invite shows no server name, README docs links 404, delegation guide dated February 2025 calls SN9 Training
Miners●●●●● 2.5
Can I compete, and what wins? · rank 29 of 32 for this audience
Q1What it is●●●●● 2
Miners compete to process as many activations as possible in the training stage.
five README bullets, no input or output shape, the linked miner docs return Page Not Found
Q2Who it is for●●●●● 2
Cuda GPU with >= 16GB VRAM (RTX 4090, for example)
and Ubuntu 22.04, no hardware table, registration cost not stated (chain burn 0.0005 TAO today), no competitiveness statement
Q3How it resists gaming or fails●●●●● 4
scoring in prose, share of run score times run allocation and burn factor, rolling window of about a day, kick policy for the bottom of the pack or below the bar, no example numbers
Q4Identity and documentation●●●●● 2
README says run setup.sh then start_miner.sh and links docs that return not found, last release v3.0.0 dated 2026-03-09, repo pushed 2026-09-08, docs folder holds design notes not a miner guide
Buyers and enterprises●●●●● 1.8
Can I use this today? · rank 23 of 32 for this audience
Q1What it is●●●●● 2
Train at Home By downloading an app, connecting your wallet and starting training
is a supply side app, no product, checkpoint or API for buyers, no statement that there is no product yet
Q2Who it is for●●●●● 1
a competitive economic alternative to training in a centralised context
is the only hint at a buyer, no customer type and no pricing anywhere
Q3How it resists gaming or fails●●●●● 2
please message us in the Bittensor Discord channel for subnet 9, or our own Macrocosmos Discord
no status page, no SLA, Discord is the only support path
Q4Identity and documentation●●●●● 2
the only documented API is the read-only auditor endpoint for compute provisioning decisions, hello@macrocosmos.ai on chain matches the site footer, no product API reference
Newcomers●●●●● 2.8
What is this and why does it matter? · rank 26 of 32 for this audience
Q1What it is●●●●● 3
The core objective of Project Orion is to show that distributed, heterogenous, permissionless training of models can be a competitive economic alternative
front page, jargon unexplained
Q2Who it is for●●●●● 3
a distributed AI ecosystem powered by people, not datacenters
and the front page economic alternative claim, no cost, ownership or coverage comparison with numbers
Q3How it resists gaming or fails●●●●● 2
the dashboard is live but every counter reads zero, results are claimed in an arXiv paper dated July 2025 and substack posts, no dated benchmark on the site
Q4Identity and documentation●●●●● 3
subnet_name iota matches the site title IOTA, url resolves, description Bringing liquid training to the world is a fragment that appears nowhere on the site
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 a 2026-09-09, rubric v1.0merged 2026-09-09links verified 2026-09-09: 37 live, 1 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.