Miami-based AI startup Subquadratic emerged from stealth Tuesday claiming to have built the first large language model that fully escapes the quadratic scaling constraint that has limited major AI systems since 2017. The company says its SubQ 1M-Preview model uses a "subquadratic architecture" where compute grows linearly with context length, potentially reducing attention compute by nearly 1,000 times compared to other frontier models at 12 million tokens.

The startup has raised $29 million in seed funding at a reported $500 million valuation from investors including Tinder co-founder Justin Mateen and former SoftBank Vision Fund partner Javier Villamizar. Subquadratic is launching three products in private beta: an API, a coding agent called SubQ Code, and a search tool called SubQ Search. The company targets a 50-million-token context window by Q4.

Subquadratic's approach, called Subquadratic Sparse Attention, selects which token-to-token comparisons matter rather than comparing every token to every other token. The company reports competitive benchmark scores, including 81.8% on SWE-Bench Verified compared to Claude Opus 4.6's 80.8%, and 95% on RULER at 128,000 tokens versus Claude Opus 4.6's 94.8%. However, the benchmarks focus narrowly on long-context retrieval and coding tasks.

The AI research community has responded with skepticism and curiosity. Some researchers question whether the claims represent a genuine breakthrough or follow the pattern of previous companies like Magic.dev, which made similar 1,000x efficiency claims in 2024 but has shown limited public progress since. Critics note the benchmark selection is narrow, with each model run only once due to high inference costs.

CEO Justin Dangel is a five-time founder, while CTO Alexander Whedon previously worked at Meta. The team includes 11 PhD researchers from major tech companies and universities. However, neither co-founder has published foundational AI research, and the company has not yet released peer-reviewed papers. The technical report is listed as "coming soon."

The fundamental challenge Subquadratic addresses is real: current transformer-based models require quadratic compute scaling, meaning doubling input size quadruples processing costs. This constraint has shaped the entire AI industry's approach to long-context processing. Independent verification of the company's claims will be crucial to determining whether this represents a genuine breakthrough in AI efficiency.