SubQ

#2 v LLM inference a serving

subquadratic · od 2026-05-05 · 16× · naposledy 30. 6. 2026

38

Momentum

SubQ (SubQ 1M-Preview) is the first commercial large language model by Miami-based Subquadratic, built on a fully sub-quadratic architecture called Subquadratic Sparse Attention (SSA). SSA replaces O(n²) dense attention with content-dependent dynamic token selection, achieving compute that scales linearly with context length. The production version offers a 1-million-token context window; the research model has been tested at up to 12 million tokens. At launch on May 5, 2026, three products entered private beta: SubQ API, SubQ Code (CLI coding agent), and SubQ Search.

Zdroje (16)