Train your own models.
Bring the training script you already have. ιsdk runs it on GPUs in many places as one run, and keeps it training when a machine leaves.
- 01
You define the workload.
Your model, your data, your objective, in the setup you already use. IOTA handles distribution, execution and recovery.
- 02
Many machines, one run.
Mixed GPUs in different buildings train one model together. When a machine leaves, its share of the work moves and the run keeps going.
- 03
Every number has a date.
Orion-100B: 100 billion parameters on 48 A100s in five US datacenters, 1 June 2026. Orion-16B: trained on mixed GPUs, machines leaving throughout. Each run is on the register.
What it cannot do yet
Per step it is slower than the same GPUs in one room: Orion-100B ran at about 65 percent of that speed. There is no price yet. Early access opens in October.
A concept by Mikyö Clark, not Macrocosmos's page
The wordmark · Back to the SDK launch