Orion-16B

#3 v AI trénovací hardware

Iota · 16b · 4× · naposledy 24. 8. 2026

82

Momentum

Orion-16B is a live AI training project by Macrocosmos based on the IOTA architecture (Bittensor Subnet 9). It is a 16-billion-parameter language model trained decentrally across three continents on a heterogeneous, permissionless pool of consumer GPUs (RTX 4090 and RTX 5090) that can scale up to 256 GPUs. Unlike its predecessor Orion-100B (fixed, verified A100 machine fleet), Orion-16B uses an open, unpredictable contribution pool without central infrastructure or ownership. The goal is to demonstrate that distributed, heterogeneous training over the public internet can be an economically competitive alternative to centralized data center training.

Zdroje (4)