Subquadratic
Research-stage sub-quadratic LLM built for multi-million-token reasoning across full codebases and long agent histories.
What it does
Subquadratic (SubQ) is developing an LLM built on a sparse-attention architecture designed to reason over context windows of up to 12 million tokens at far lower compute cost than standard transformers. It offers an API for developers and a coding-agent integration for tools like Claude Code, Codex and Cursor.
Core features
Up to 12M-token context window
Sub-quadratic sparse-attention architecture (claimed ~1000x less attention compute at scale)
OpenAI-compatible API with streaming and tool use
Coding-agent plugin that offloads token-heavy context work
Benchmarked long-context retrieval and reasoning performance
Best for
→Reasoning across an entire code repository in one call
→Maintaining long-running agent state without quality loss
→Reducing compute cost for long-context AI applications
→Speeding up coding agents on large codebases
Toolspool rankingby global site rank