SubQ: First Fully Sub-Quadratic LLM

Author: @subquadratic + @alex_whedon | Date: 2026-05-05

SubQ is announced as the first fully sub-quadratic LLM achieving 12M-token context windows at 150 tokens per second. This represents a major breakthrough in efficient long-context processing, bypassing the quadratic attention bottleneck entirely.

Key Claims

  • First LLM to achieve fully sub-quadratic complexity (not just approximate)
  • Supports 12 million token context length
  • Generates at 150 tokens/second despite massive context
  • Eliminates the fundamental O(n²) attention scaling barrier

Takeaways

  • Long-context architectures are moving beyond approximations to fundamentally different compute patterns
  • 12M tokens opens up repository-scale code understanding, book-length reasoning, and multi-document synthesis
  • Speed remains practical despite extreme context lengths

Linked Concepts