Open-Source 27B Model Cluster Rivals 1.6T AI Giant at Far Lower Cost
A new open-source multi-model system called Fusion-MOA combines several 20B–30B parameter models to match or outperform much larger commercial AI flagships on real engineering tasks. The system uses a collaborative architecture where multiple models anonymously cross-review each other's answers, enabling self-correction on problems that stumped each model individually. On benchmark tests, the cluster matched a 744B-class model and outperformed a 1.6-trillion-parameter cloud system by 15 percentage points. Technically, it achieves decoding speeds of up to 62 tokens per second, supports a 128K context window, and ran 906 consecutive calls over an hour without a single restart. The system is designed to run on domestic and consumer-grade hardware without relying on export-restricted chips, making it a cost-effective alternative to metered cloud AI services.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
Discussion (0)
Log in to join the discussion and vote.
Log in