What is IOTA Train at Home? How decentralized training on SN9 works

If you've just installed Train at Home, or you're wondering whether it's worth running on a spare Mac, here's the short version: your Mac helps train a large AI model that no single computer could train alone, and in return you can earn rewards on the Bittensor network. This article explains what's actually happening behind that sentence.

IOTA, in one paragraph

IOTA (Incentivised Orchestrated Training Architecture) is a project by Macrocosmos that runs on Bittensor Subnet 9 (SN9). According to Macrocosmos's own documentation, IOTA is "a data- and pipeline-parallel training algorithm designed to operate on a network of heterogeneous, unreliable devices in adversarial and trustless environments." In plain language: it lets a large number of very different, sometimes-offline computers — instead of one big datacenter of matched, reserved hardware — cooperate to pre-train a large model together.

Train at Home is the consumer-friendly way into that network: a desktop app that lets you "plug in any hardware you have access to and earn rewards for powering decentralized AI model training," as Macrocosmos's own product page puts it, without needing to run a full validator or miner node by hand.

Why IOTA exists: the problem it fixes

IOTA didn't appear out of nowhere. Subnet 9 existed before it, and — per Macrocosmos's own subnet documentation — already proved something important in August 2024: "a distributed network of incentivized, permissionless actors could each pre-train large language models" competitively. But that first version had two structural problems the documentation itself names directly:

• Every miner had to fit an entire model locally. If your hardware couldn't hold the whole model, you couldn't meaningfully participate.
• "Winner-takes-all" rewards encouraged model hoarding. Because miners competed as isolated rivals rather than collaborators, there was little incentive to share progress.

IOTA's answer is to turn "SN9's previously isolated competitors into a single cooperating unit that can scale arbitrarily while still rewarding each contributor fairly." Concretely, that means splitting one large model into layers spread across many machines, instead of asking every machine to hold the whole thing.

How the training actually flows

Based on IOTA's own technical primer, the system is built around a hub-and-spoke shape, centered on an orchestrator:

• The model is split into layers. Your Mac, at any given time, is typically responsible for one layer, not the whole model.
• The orchestrator coordinates who works on what, and hands out the pieces of data (activations) that need to move from one layer to the next.
• Those activations move between layers over shared object storage (Cloudflare R2 — confirmed by the presigned upload URLs the app itself requests) and, when a direct path exists, a peer-to-peer transport (a QUIC-based technology called iroh). Based on the technical primer's architecture description together with what the app's own logs show (a miner with no reachable P2P peers still keeps training), object storage appears to be the baseline path and P2P an optimization layered on top of it — the primer doesn't spell out that exact relationship, so treat it as our best-supported reading rather than a confirmed fallback mechanism.
• Your Mac reports back to the orchestrator through periodic heartbeats, which include the current run, epoch and phase — things like training, weights_uploading, or merging_partitions. This is also exactly what tools like subnera read to show you a real status instead of a frozen screen.

This design is explicitly meant to tolerate machines that are slow, that come and go, or that sit behind restrictive home networks — a home Mac was never expected to behave like a datacenter node, and the whole point of routing most data through shared storage rather than requiring every machine to talk directly to every other machine is to make that tolerable.

Why bandwidth isn't the bottleneck you'd expect

A natural worry: "my home internet can't possibly keep up with training a large model." IOTA's own technical primer addresses this directly — it describes compression techniques that cut the amount of data that needs to move between layers by roughly 100–300× compared to raw activations, explicitly framing "typical" home internet bandwidth (50–200 Mbps) as the target it's engineered for, not an edge case it merely tolerates. That's also why the official guidance about your connection stays qualitative ("keep a stable network") rather than demanding a specific speed — see the optimal setup guide for what that means in practice.

What this means for rewards

Macrocosmos's subnet documentation describes the incentive design only at a high level — it "aligns participants through transparent token economics" while "rewarding each contributor fairly" rather than the old winner-takes-all model. The precise mechanics of how much a given contribution earns aren't detailed in the material this article draws from, so we won't guess at numbers here; the rewards and payouts guide covers how to track what you actually receive.

What your Mac is (and isn't) doing

To be concrete about scope: your Mac is a worker that trains one layer at a time, exchanges activations through storage (and sometimes directly with peers), and reports its status upward. It is not running a full copy of the model, not acting as a public server for anyone else's traffic beyond what the orchestrator assigns it, and not exposed to the internet in a way that requires you to open ports — see the network and ports guide for exactly what does and doesn't need attention on your end.

If you haven't set your Mac up yet, or want to make sure it's contributing as steadily as possible, the optimal setup guide is the practical next step: install, keep-awake, and the handful of checks worth knowing about.