Subquadratic, a Miami-based startup, came out of stealth on May 5, 2026 with $29 million in seed funding and a large language model, SubQ 1M-Preview, that it says is the first frontier LLM built without the standard transformer attention mechanism. The launch drew immediate scrutiny from AI researchers who said the company's headline efficiency claims cannot yet be independently checked.
What Subquadratic shipped
Most large language models rely on dense attention, where compute and memory scale quadratically with the length of the input. Subquadratic built what it calls Sparse Selective Attention (SSA), a mechanism that uses learned, content-aware sparsity instead of the fixed patterns other efficient-attention approaches have tried. The company says this brings scaling down to roughly O(n log n) or O(n), rather than O(n squared), and supports a 12 million token context window.
Subquadratic's own comparisons put SSA at 52 times faster than FlashAttention, with 63% less compute at the architecture level, and claims a roughly 1,000x reduction in attention compute at long context compared to dense-attention frontier models. Those are the company's own figures from its launch materials, not third-party measurements.
The $29 million seed round valued the company at a reported $500 million. Investors include Tinder co-founder Justin Mateen, former SoftBank Vision Fund partner Javier Villamizar, and early backers of Anthropic, OpenAI, Stripe and Brex, according to Pulse2 and Refresh Miami. CEO Justin Dangel, a five-time founder with a background in health tech, insurance tech and consumer products, co-founded the company with Alex Whedon.
Why it matters for buyers, not just researchers
If SSA holds up, it targets a real cost problem for anyone running long-document workloads: legal discovery, codebase-wide reasoning, multi-document retrieval, or agent systems that need to hold large working context without re-paying quadratic attention costs on every pass. A model that can genuinely serve a 12 million token window at a fraction of dense-attention compute would change the cost math for those workloads specifically, not for chat-style short-context use cases where the difference matters less. That is the pitch enterprise buyers evaluating inference spend will want to test for themselves before it changes any procurement decision.
The pushback, and why it happened fast
Within hours of the launch, the response split into two camps. AI commentator Dan McAteer summarized the mood in a widely shared post: "SubQ is either the biggest breakthrough since the Transformer... or it's AI Theranos," as reported by VentureBeat.
The skepticism centers on three specific gaps. No weights or full technical report accompanied the launch, so independent verification of the architecture's actual behavior is not possible without either. On benchmark methodology, The New Stack's reporting says each benchmark was run only once due to inference cost, with no confidence intervals reported, and the company's own paper acknowledges its SWE-Bench margin is partly a function of the test harness rather than the model alone. And on the MRCR v2 long-context benchmark, Subquadratic reported a research score of 83, while the production model that shipped scored 65.9 in third-party testing, a 17-point drop the company has not explained.
Dangel addressed the skepticism directly rather than dismissing it: "Extraordinary claims will often be greeted rightly with skepticism. The fact that our company has a potentially industry-disrupting innovation, I'm not surprised by the reaction."
What to watch
The open questions are concrete and checkable, which is unusual for an AI architecture claim this large. Independent researchers will want the technical report and, ideally, open weights or API access broad enough to run the standard long-context benchmark suites themselves, with multiple runs and reported variance rather than single-shot numbers. The 17-point MRCR gap between the lab result and the shipped model is the single most testable discrepancy, and whether Subquadratic can account for it will do more to settle the debate than any efficiency multiplier in its launch deck.
For now, SubQ is a funded, shipped product with a real context window enterprises can test against their own workloads, and a set of efficiency claims that remain the company's own numbers until someone outside Subquadratic reproduces them.
Sources: VentureBeat: Miami startup Subquadratic claims 1,000x AI efficiency gain · Pulse2: Subquadratic $29 Million Seed Raised · Refresh Miami: Subquadratic raised $29M · Subquadratic: Introducing SubQ