Skip to content
MIT Technology Review · AI

Subquadratic shares third-party benchmarks for its SubQ model, claiming it breaks the quadratic attention bottleneck

A startup claims it broke through a bottleneck that’s holding back LLMs

Miami startup Subquadratic now shares independent benchmarks from Appen for its SubQ model, which it claims solves the quadratic attention bottleneck that makes LLMs slow and expensive. SubQ can process up to 12× more text at once than most models, with faster speed and lower cost, while roughly matching top models from DeepMind, OpenAI, and Anthropic on coding tasks. The CTO admits they should have released third-party results alongside the initial announcement. Appen's director of generative AI research calls the results exciting and validating. SubQ is not yet publicly available, and the post does not disclose specific latency, power, or pricing figures.

Why it matters: Subquadratic claimed its SubQ model broke the quadratic cost bottleneck of dense attention without evidence, and has now released third-party benchmark numbers that give the dispute testable data.

Read the original ↗Export Markdown