Woofun AI reports that Subquadratic has published a technical report for SubQ version 1.1, addressing prior criticism regarding the absence of independent validation. In collaboration with assessment firm Appen, the company claims the model achieves 98% retrieval accuracy within a 12 million token context window and performs comparably to state-of-the-art models in programming tasks. The report discloses that the model utilizes incremental training on 1 trillion tokens atop an open-source foundation, replacing standard attention calculations rather than training from scratch.

Despite these assertions, the developer community remains skeptical, arguing that the update lacks fundamental technological innovation. Critics note that the system primarily applies established block-sparse attention mechanisms by segmenting long texts for dynamic filtering.

Furthermore, concerns were raised about AI-generated clichés in the documentation and potential scheduling overhead that could cause severe latency for the slowest 1% of users during concurrent multi-user access. As SubQ has not released core parameters or an open API, its claims of reduced computational costs remain unverified.