Miami's Subquadratic Launches With $29M and a 12M-Token LLM That Claims to Cut Compute 1,000x

Miami-based startup Subquadratic emerged from stealth on May 5 with $29 million in seed funding and a commercial large language model that claims to do something no other production system can: hold 12 million tokens in context — roughly 9 million words, or 120 paperback novels — without the quadratic compute blow-up that hobbles transformer attention. The company calls the model SubQ, and its founders are pitching it as the first frontier-class LLM to break the scaling rules that OpenAI, Anthropic, and Google DeepMind have all built their stacks on.

Co-founders Justin Dangel (CEO) and Alexander Whedon (CTO) built SubQ on a proprietary technique they call Subquadratic Sparse Attention, or SSA. Where standard dense attention compares every token to every other token — work that grows as the square of the context length — SSA picks only the most relevant tokens for each comparison, effectively turning quadratic scaling into roughly linear scaling. Dangel framed the architectural shift bluntly: "The fundamental scaling laws imposed by the transformer architecture and dense attention have been broken through."

The performance claims are extreme even by frontier-AI standards. At 1 million tokens of context, Subquadratic says SubQ is 50 times faster and 50 times cheaper than competing models while maintaining higher accuracy. At the full 12-million-token context window, the company reports a roughly 1,000x reduction in compute. On the long-context RULER 128K benchmark, SubQ reportedly hits 95 percent accuracy for about $8 per evaluation, compared to 94 percent at roughly $2,600 for Anthropic's Claude Opus — a 300x cost gap if the numbers hold.

The seed round was led by former SoftBank Vision Fund partner Javier Villamizar and Tinder co-founder Justin Mateen's JAM Fund, with participation from early backers of Anthropic, OpenAI, Stripe, and Brex. Subquadratic is launching three products alongside the model: a developer API, a CLI coding agent called SubQ Code that ingests whole codebases at once, and a free consumer search tool. Independent researchers have pushed back hard on the 1,000x number, demanding third-party benchmarks before treating it as anything more than marketing — but if even a fraction of the speed-up survives outside testing, the transformer's quadratic ceiling stops being a law of nature.