Miami-based AI startup Subquadratic has emerged from stealth mode claiming to have resolved a decade-old computational bottleneck in large language models (LLMs). Initial skepticism gave way after independent tests validated key aspects of its novel SubQ model, which processes significantly more text at lower cost and energy use while maintaining competitive performance.
- SubQ processes up to 12 times more text than current LLMs
- Independent tests confirm lower cost and energy use
- Potential to move beyond transformer architectures
What happened
Subquadratic, a Miami-based AI startup, announced it developed a new large language model called SubQ that claims to overcome a longstanding computational bottleneck associated with transformers, the dominant architecture behind today’s LLMs. The company asserts that SubQ is significantly faster, cheaper, and more energy-efficient compared to leading models from Google DeepMind, OpenAI, and Anthropic.
Initially, Subquadratic faced skepticism due to limited independent evidence supporting their claims. However, recent third-party benchmarking by Appen has substantiated many of SubQ’s performance improvements. These results highlight SubQ's ability to handle up to 12 times more text simultaneously and maintain comparable accuracy on complex tasks such as coding and document analysis.
Why it matters
Current large language models rely heavily on a process called dense attention within transformer networks, which involves a computationally expensive operation growing quadratically with input length. This limits the amount of text these models can efficiently process and drives high power consumption, making scaling costly and environmentally taxing.
Subquadratic’s approach potentially breaks this quadratic barrier, enabling models to analyze much larger text corpora at lower cost and energy demand. This breakthrough could unlock new possibilities for AI applications requiring extensive context, such as comprehensive codebase analysis or multi-document synthesis, while reducing operational barriers in AI deployment.
What to watch next
Subquadratic plans to continue validating its SubQ model through further independent trials and intends to release more detailed information to the AI research community. Industry observers will be looking for wider access to SubQ to verify its real-world performance and practical impact across diverse AI workloads.
Longer term, the company hopes SubQ will initiate a shift away from traditional transformer models, which have dominated since 2017, toward a new era of more efficient large language models. The success of this transition could influence future AI hardware design, software frameworks, and the economic landscape of AI service providers.