Miami AI Startup Shatters Context Window Records with 12M-Token Model

A Miami-based AI startup has released a model with a record-breaking context window. Subquadratic’s first product can process up to 12 million tokens—far beyond what most competitors offer. The company claims its technology also runs faster and more efficiently than traditional designs. Subquadratic’s new model uses an architecture called Subquadratic Selective Attention (SSA). Unlike standard systems, this approach scales linearly in compute and memory as context length grows. Tests show it operates 52 times faster than dense attention at a million tokens.

Performance benchmarks place the model ahead of rivals in key areas. It scored 82.4% on SWE-bench, outperforming Anthropic’s Opus 4.6 and Google’s Gemini 3.1 Pro. On the multi-reference retrieval test MRCR v2, it achieved 83, surpassing GPT-5.5’s 74.0%. The company has already secured $29 million in funding and holds a $500 million valuation. Two products are now in beta: an API for the 12M-token window and SubQ Code, a command-line agent. A third tool, SubQ Search, will also be available. Looking ahead, Subquadratic plans to expand further, with a 50-million-token model in development. Industry trends suggest that by 2026, most frontier models will advertise context windows of at least a million tokens. However, few are expected to use that capacity effectively—highlighting the potential advantage of Subquadratic’s approach.

Subquadratic’s model sets a new standard for context length and speed. The startup’s funding and valuation reflect strong investor confidence in its technology. With plans for even larger models and commercial tools, the company aims to reshape how AI processes long-form information.