Subquadratic publishes Appen benchmarks for sparse AI model

Miami-based artificial intelligence startup Subquadratic has published benchmark results it says support its claim that a proprietary language-model architecture can handle much larger text inputs while using less computational energy. The evaluation was carried out by testing firm Appen, according to the company.

Subquadratic said the Appen benchmarks indicate its model can process text volumes 12 times larger than standard architectures. The company framed the release as independent validation after its earlier technical claims drew scrutiny across the industry.

Appen’s evaluation and what it verified

Appen Director of Generative AI Research Jeanine Sinanan-Singh Appen Director of Generative AI Research Jeanine Sinanan-Singh said the tests validated the architecture’s ability to address speed and efficiency constraints found in existing systems. Subquadratic presented the results as evidence that its approach can reduce the computational burden associated with long-context processing. The company’s announcement arrives after skepticism followed initial claims last month that it had resolved a mathematical bottleneck in standard transformer models. Engineer Dan McAteer had described the then-unverified model as either a major breakthrough or the artificial intelligence equivalent of Theranos, reflecting the level of uncertainty before third-party measurement. Why long text becomes expensive in standard transformer models Subquadratic pointed to a core limitation in widely used large language model designs: dense attention inside transformer architectures. In that setup, when text is processed, each token is encoded and combined with every other token in the sequence, creating rapidly rising computational requirements as inputs grow.

Subquadratic Chief Executive Justin Dangel said that tasks

The company illustrated the scale of that cost with a specific example: a 10,000-word text can require nearly 50 million individual multiplications. Subquadratic Chief Executive Justin Dangel said that tasks such as summarizing an entire book require cross-referencing early and late parts of the text alongside many other combinations, increasing the amount of work the model must do.

In the company’s description, the constraint follows a quadratic pattern: doubling the word count roughly quadruples the number of required computations. Subquadratic said this “quadratic expansion” limits how much data models can ingest in one pass, shaping both performance and cost.

Sparse attention approach and enterprise positioning

Subquadratic said its SubQ architecture replaces dense computation with sparse attention to lower calculation requirements. The company also said that, in specific coding tasks, the model maintains performance parity with leading platforms from Google DeepMind, OpenAI, and Anthropic.

Subquadratic Chief Technology Officer Alex Whedon said publishing the Appen benchmarks was intended to preempt further questions about the system’s technical capabilities. The company noted the model is not yet available for public testing.

Subquadratic said its target users are enterprises with workloads that require simultaneous analysis of very large code bases or extensive document repositories. For now, the benchmarks are positioned as the primary external reference point for the model’s claimed efficiency gains, while broader access and additional verification remain unresolved.