Miami-based AI startup Subquadratic says its SubQ model replaces dense attention, the core operation in transformer-based LLMs, with a sparse-attention mechanism that dynamically chooses which token relationships to compute. The company says the design makes the model faster, cheaper and more energy efficient, while allowing it to process up to 12 times as much text at once as most other models.
Third-party evaluator Appen tested SubQ and reported results that support some of the company’s claims. In a speed benchmark, SubQ was 56 times faster than models using FlashAttention, and it scored 89.7% on LiveCodeBench, putting it near leading coding models. Appen also said SubQ scored 98% on a needle-in-a-haystack retrieval test with context windows six million and 12 million tokens long.
Subquadratic is positioning SubQ for coding and large-scale data search, but the model is not yet widely available. The company says tens of thousands of potential users have signed up for early access, including more than 500 enterprise customers. Skepticism remains because benchmarks offer a limited view of real-world performance, and SubQ reused weights from a version of the open-source Qwen model rather than being trained entirely from scratch.
