MIAMI — Subquadratic, an artificial intelligence startup based in Miami, announced its exit from stealth mode last month. The company shared results from an independent evaluation of its large language model, SubQ, conducted by the third-party firm Appen.

Subquadratic developed SubQ and claims the model can process up to 12 times as much text at once compared to most other models. The company also claims SubQ matches the performance of models from Google DeepMind, OpenAI, and Anthropic on key coding tasks. Appen found that SubQ was 56 times faster than models using FlashAttention in a baseline speed test and scored 89.7% on LiveCodeBench, a test that assesses performance on competitive coding problems.

SubQ has a context window of up to 12 million tokens, while most top large language models currently have context windows of one million tokens. Subquadratic uses sparse attention in its model architecture, dynamically selecting which words to focus on during processing. Alex Whedon, chief technology officer at Subquadratic, stated that historical mechanisms used fixed patterns. "Historically, most mechanisms have used fixed patterns, like always comparing the first word to the fifth," Whedon said. "That's pretty limiting. Language is too sophisticated for that." He added, "Sparse attention says not all of those relationships are important, because they're not."

Justin Dangel, cofounder and CEO of Subquadratic, said he hoped the company is initiating a new age of efficiency. "We hope we're kicking off a new age of efficiency," Dangel said. "We don't think anybody will be building on transformers in a few years." Subquadratic claims it costs eight dollars to run SubQ through the RULER 128 test, a benchmark developed by Nvidia to assess a model's ability to retrieve information from large datasets. Dangel stated that running Anthropic's LLM Opus 4.6 through the same test costs $2600.

Jeanine Sinanan-Singh, Appen's director of generative AI research, said the results validated Subquadratic's architecture. "That was really exciting to me, it validated their architecture," Sinanan-Singh said. "I was like, 'Wow, this could be a game changer,' because models struggle with speed and inefficiency." She noted, "But when you have kind of shocking results, it's really not as credible when you say it yourself." Whedon acknowledged the initial skepticism surrounding Subquadratic's claims. "We expected healthy skepticism," Whedon said. "In hindsight, releasing the third-party benchmarks alongside the initial announcement would have preempted much of the skepticism, which is why we're taking the time to make sure any future results are fully verified before putting them out."