Quick Facts
- Subquadratic claims its SubQ model delivers 1,000x efficiency gains over frontier AI models at 12 million token contexts
- The Miami startup raised $29 million in seed funding with a team of 11 PhD researchers from Meta, Google, and other tech giants
- AI researchers demand independent verification, with some comparing the claims to Theranos due to lack of technical transparency
Miami-based AI startup Subquadratic has sparked intense debate in the machine learning community with claims that its SubQ model delivers efficiency gains of up to 1,000 times over existing frontier AI models.
The company, led by CEO Justin Dangel and CTO Alexander Whedon (former Head of Generative AI at Meta), claims SubQ can process 12 million token contexts while being 50 times faster and cheaper than leading models. The startup raised $29 million in seed funding from investors including Tinder co-founder Justin Mateen and former SoftBank Vision Fund partner Javier Villamizar.
SubQ uses what Subquadratic calls “Subquadratic Selective Attention,” an architecture that identifies which token comparisons matter rather than comparing every token to every other token. On the RULER 128K benchmark, SubQ scored 95% accuracy at $8 cost, compared to Claude Opus’s 94% accuracy at $2,600 cost.
However, prominent AI researchers have expressed skepticism about the claims. AI commentator Dan McAteer wrote that “SubQ is either the biggest breakthrough since the Transformer… or it’s AI Theranos.” The comparison to the infamous blood-testing fraud reflects the scale of Subquadratic’s assertions.
AI engineer Will Depue noted that SubQ is “almost surely a sparse attention finetune of Kimi or DeepSeek.” CTO Whedon confirmed on X that the company uses “weights from open-source models as a starting point.”
The skepticism stems partly from previous failed attempts at subquadratic architectures. Models like Mamba, RWKV, and Hyena demonstrated linear scaling but underperformed dense attention models on downstream benchmarks at frontier scale.
Critics also question why Subquadratic gates access through an early-access program if the model truly costs less than 5% of existing options. Developer Stepan Goncharov called the benchmarks “very interesting cherry-picked benchmarks.”
The situation echoes Magic.dev’s August 2024 announcement of a 100-million-token model with claimed 1,000x efficiency gains. Magic raised roughly $500 million based on those claims, but there is no public evidence of the model being used outside Magic as of early 2026.
Subquadratic plans to launch three products: SubQ API for developers, SubQ Code for loading entire codebases into context, and SubQ Search as a free consumer research tool to compete with Perplexity and ChatGPT.
The company promises to release a technical report, model weights, and allow independent benchmarks. CEO Dangel said the team is “very focused on the problem of how we transition from a dense attention, quadratic scaling architecture to a sparse attention linear architecture.”
AI researcher John Rysana defended the work, calling it “subquadratic attention done well” and saying “odds of it being BS are extremely low.” However, the machine learning community remains divided until independent verification becomes available.
This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.