The context window has been shattered: Subquadratic debuts a 12-million-token window

Miami startup Subquadratic launched its first model with a 12-million-token context window, powered by its new Subquadratic Selective Attention architecture.

The company says the architecture scales linearly and is more efficient than dense attention, with plans to expand to 50 million tokens soon.

Subquadratic claims strong results on retrieval and coding benchmarks, including wins over several leading models on MRCR v2 and SWE-bench.

Access is available now through an API, plus productized tools called SubQ Code and SubQ Search.