Miami-based AI startup Subquadratic emerged from stealth on May 5, 2026, unveiling its first large language model, SubQ, alongside a $29 million seed round. The company says SubQ can process up to 12 million tokens in a single prompt, equivalent to roughly 9 million words or about 120 books. Most leading frontier models currently cap out at 1 million tokens of context.
SubQ is built on a new architecture the company calls Subquadratic Sparse Attention, or SSA. In conventional transformer models, every token is compared with every other token, so attention cost scales quadratically with sequence length. Subquadratic says SSA instead uses content-dependent selection to compute attention only over the relationships that actually matter, allowing cost to grow linearly with context length. The company reports 95.0% accuracy on RULER 128K, 81.8% on SWE-Bench Verified, and 65.9% on MRCR v2 at 1 million tokens.
Subquadratic was founded by CEO Justin Dangel and CTO Alexander Whedon, and employs 11 PhD researchers. Investors include Tinder co-founder Justin Mateen, former SoftBank Vision Fund partner Javier Villamizar, and early backers of Anthropic, OpenAI, Stripe, and Brex. Reports place the company's valuation at roughly $500 million. Three products are entering private beta: a SubQ API exposing the full context window, a CLI coding agent called SubQ Code, and a long-context research tool named SubQ Search that the company plans to offer free of charge.
The company claims SSA reduces attention compute by nearly 1,000x at 12 million tokens compared with dense-attention transformers, but parts of the research community remain skeptical. Earlier subquadratic approaches, including Mamba, RWKV, DeepSeek Sparse Attention, and Kimi Linear, have generally underperformed dense attention at frontier scale. Observers also point to Magic.dev, which raised around $500 million in 2024 on a 100-million-token claim that has yielded little visible production deployment. Subquadratic says a 50-million-token model is targeted for the fourth quarter, and that while weights will not be open-sourced, enterprise customers will be able to fine-tune SubQ on their own data.