Subquadratic's new SubQ model claims to process 12x more context than transformers
subquadratic released third-party benchmarks for its subq llm, claiming it offers 12x the context window of transformer models at a fraction of the cost, threatening to upend standard architectures.
subquadratic released third-party validation from appen supporting its claim that subq solves the mathematical bottleneck of transformer dense attention, enabling models to process up to 12 times more context at a fraction of the compute cost. the benchmarks indicate that subq maintains competitive performance against leading models from google deepmind and openai on specific tasks like coding, while significantly boosting efficiency. if scalable, this architecture could shift the industry away from transformers, which have dominated since 2017, by offering a viable path for analyzing massive datasets like entire legal documents or codebases without the usual computational drag.