Miami Startup SubQ Claims to Break AI’s Decade-Old Speed Bottleneck-Is It Really a Game-Changer?

Subquadratic, a Miami-based startup, claims to have bypassed the quadratic attention bottleneck in large language models (LLMs) with its SubQ model. According to independent tests by Appen, SubQ operates 56 times faster than models using FlashAttention and maintains near-perfect retrieval across context windows of up to 12 million tokens.

The company emerged from stealth mode in May 2026, announcing a new architecture it says is faster, cheaper, and more energy-efficient than existing models from OpenAI, Google DeepMind, and Anthropic. While initial claims were met with skepticism due to a lack of evidence, Subquadratic has since released third-party evaluation results to support its assertions.

SubQ can process up to 12 times more text simultaneously than most current models. This capability allows the model to analyze entire code bases or hundreds of documents at once, according to the company.

The startup says SubQ matches the performance of frontier models on key tasks like coding while significantly reducing the computational load. This is achieved by replacing the standard “dense attention” mechanism used in transformers with a “sparse attention” approach.

Most LLMs rely on transformers that multiply every token in a text chunk by every other token to determine meaning. This process creates a quadratic expansion of computation; doubling the text length roughly quadruples the required processing power, according to Subquadratic CEO Justin Dangel.

SubQ uses sparse attention to select only the most relevant tokens for multiplication rather than processing every possible relationship. CTO Alex Whedon said the mechanism is unique because it dynamically selects which tokens are important on the fly, rather than using fixed patterns.

Subquadratic describes this dynamic selection as the “secret sauce” of the architecture, though the company has not disclosed the specific method used to choose which words the model focuses on.

What do independent benchmarks show for SubQ?

Third-party firm Appen conducted evaluations to verify Subquadratic’s claims. In theoretical speed tests, Appen found SubQ was 56 times faster than models utilizing FlashAttention, a previous sparse-attention technique.

On the LiveCodeBench test, which uses competitive coding problems from real contests, SubQ scored 89.7%. Jeanine Sinanan-Singh, Appen’s director of generative AI research, stated the model provides “frontier-level performance in coding.”

Appen also performed “needle-in-a-haystack” tests to measure the model’s ability to retrieve specific information from massive datasets. SubQ scored 98% with context windows of six million and 12 million tokens.

Dangel claimed the cost efficiency is equally significant. He stated that running Anthropic’s LLM Opus 4.6 through the Nvidia-developed RULER 128 test cost $2,600, while the same test cost SubQ eight dollars.

In a demonstration, Whedon used SubQ to reason across 400 documents in seconds. The same task caused Perplexity, an LLM-powered search engine, to fail to load the full document set.

Why are some researchers skeptical of the claims?

Some industry observers question whether Subquadratic has truly solved the quadratic bottleneck. Will Depue, an independent AI researcher and former OpenAI employee, noted that while the results may be useful, the current public evidence doesn’t fully justify the claim of solving the bottleneck.

One primary point of contention is that Subquadratic did not train SubQ from scratch. Instead, the company reused weights from Qwen, a Chinese open-source model, to bootstrap the system. Depue suggested this complicates the claim that the company fully reinvented how LLMs work.

Access to the model remains limited. Although the company reports more than 500 enterprise customers and tens of thousands of sign-ups, most users remain on a waitlist. Subquadratic attributed this to limited resources as a small company.

The skepticism was highlighted early on by AI engineer Dan McAteer, who posted on X that SubQ was either a massive breakthrough or “AI Theranos.”

Whedon acknowledged the reaction, stating that releasing third-party benchmarks alongside the initial May announcement would have preempted much of the doubt.

Despite the criticism, Dangel believes the industry is moving away from traditional transformers. He stated that he doesn’t believe anyone will be building on transformers in a few years.

We hope we’re kicking off a new age of efficiency. We don’t think anybody will be building on transformers in a few years.

Justin Dangel, Subquadratic CEO