Originally published on Poniak Times . Reposted here for the developer and AI engineering community. Subquadratic’s SubQ model claims to make long-context AI more efficient through sparse attention. The claim is serious, but it still requires independent validation before being treated as a major shift in AI architecture. The artificial intelligence industry has spent the past few years moving in one clear direction: larger models, larger context windows, larger GPU clusters, and larger infrastructure bills. From frontier language models to enterprise AI copilots, the common belief has been that higher capability usually requires more compute, more training data, and more expensive hardware. Subquadratic, a relatively new AI research and infrastructure company, is now challenging that assumption with a model called SubQ 1M-Preview.…