Subquadratic Claims Breakthrough in Overcoming Transformer Limitations, Backed by Independent Evaluation

Miami-based AI startup Subquadratic claims to have solved a long-standing mathematical bottleneck that has constrained the performance of large language models (LLMs) for years. The company’s newly developed model, “SubQ,” is said to introduce a novel solution to the computational complexity issues inherent in the attention mechanism of traditional Transformer architectures. While the initial unveiling of SubQ last month was met with skepticism due to a lack of detailed evidence, independent third-party testing by evaluation agency Appen has now been released, lending credence to some of the company’s claims.

The Limitations of Transformers and SubQ’s Claims

Since its introduction in 2017, the Transformer architecture has become the foundation of modern LLMs. However, the computational cost of its attention mechanism increases quadratically with the number of input tokens (O(n²)), leading to a significant slowdown in processing speed and a sharp rise in energy consumption as context length increases. This scalability issue has posed a major challenge for data-intensive tasks such as analyzing lengthy documents or processing large codebases.

Subquadratic asserts that it has managed to overcome this O(n²) limitation in a practical way. According to the company, SubQ can process up to 12 times more text simultaneously while maintaining performance that matches or exceeds the best existing models. Specifically, it claims that SubQ can analyze hundreds of documents or entire codebases at once. Furthermore, the model reportedly offers improved processing speed, reduced costs, and lower energy consumption. These features present a highly attractive proposition for companies struggling with the rising costs of cloud-based inference.

Alex Whedon, Subquadratic’s Chief Technology Officer (CTO), admitted, “We anticipated healthy skepticism,” acknowledging that the initial announcement lacked sufficient transparency. “Looking back, if we had released third-party benchmarks at the time of the announcement, much of the skepticism could have been avoided. Moving forward, we will only release verified results,” he said.

Independent Evaluation Highlights Potential

Subquadratic commissioned Appen, a third-party evaluation agency, to assess SubQ’s performance. Jeanine Sinanan-Singh, Appen’s Director of Generative AI Research, expressed enthusiasm about the results, stating, “We were thrilled to see the results. It confirmed the validity of SubQ’s architecture.” She further emphasized the importance of independent evaluations in building trust, saying, “Issues like model speed and inefficiency are challenges for the entire industry, but shocking results are more credible when demonstrated by third-party organizations rather than the company itself.”

Appen’s test results revealed that SubQ achieved scores on major LLM benchmarks that were on par with or superior to leading models from Google DeepMind, OpenAI, and Anthropic, particularly in coding-related tasks. However, Sinanan-Singh also noted, “SubQ is not a replacement for the top existing models across all tasks. Its strengths lie in delivering drastic speed improvements and cost reductions for specific tasks.”

Justin Dangel, Subquadratic’s co-founder and CEO, stated, “We aim to usher in a new era of efficiency,” sharing his long-term vision that “in a few years, no one will be building models with Transformers anymore.” The company has pledged to continue publishing verified results gradually.

Industry Reactions and Future Challenges

Subquadratic’s announcement has sparked polarized reactions. AI engineer Dan McAteer remarked on X (formerly Twitter), “SubQ is either the biggest breakthrough since Transformers or the Theranos of AI.” Such divided opinions reflect the current lack of widespread validation for the company’s claims.

As of now, SubQ is not publicly available, and external researchers and developers cannot independently test the model. While Subquadratic has announced plans to release more detailed technical documentation and an API, no specific timeline has been provided.

From a technical standpoint, one of the biggest lingering questions is how Subquadratic overcame the O(n²) limitation mathematically. If the company chooses not to disclose these details due to patents or trade secrets, academic validation and reproducibility will be challenging.

Editorial Opinion

In the short term, Subquadratic’s independent evaluation results could introduce a new dimension of competition to the LLM industry. Interest in architectures beyond Transformers is likely to grow, and established players like Google and OpenAI may accelerate their research into alternative methods. The practical implementation of SubQ could expand engineering options, particularly in production environments where reducing cloud inference costs is critical. If the company releases an API within the next three to six months, it is anticipated that pilot experiments for data-intensive tasks will commence.

From a long-term perspective, whether SubQ can genuinely end the dominance of Transformers will depend on its scalability and versatility. Thus far, its superiority has only been demonstrated in specific tasks, and its potential as a foundational model for diverse tasks remains unproven. Over the next one to three years, Subquadratic must establish its competitiveness as a large-scale pre-trained model. Without this, its claims may ultimately be categorized as a niche optimization technique. Transforming the industry will require participation in an open ecosystem rather than relying on a closed approach.

References

• MIT Technology Review AI — Published on 2026-06-19