Subquadratic (SubQ)

Overview

SubQ is the first large language model built on a fully sub-quadratic sparse-attention (SSA) architecture, designed specifically for long-context reasoning tasks. Unlike transformer-based models that scale attention compute at O(n²), SubQ scales at O(n), reducing attention compute by nearly 1,000× at 12M tokens.

SubQ targets developers, enterprise teams, and coding agents that need to reason across entire codebases, months of pull request history, and long-running agent state — all in a single prompt call at a fraction of the cost of comparable frontier models.

Key Benefits

• 12M token context window allows reasoning across full repositories and extended agent histories in one prompt.
• Runs at 150 tokens per second with costs at one-fifth of other leading LLMs.
• Scores 95.6% on RULER @ 128K, outperforming Claude Opus 4.6 (94.8%) on long-context accuracy.
• Scores 86.2% on MRCR v2 (8-needle, 1M), surpassing all listed competitors including Claude Opus 4.7 (74.0%).
• All benchmark results are third-party validated.
• SubQ Code delivers approximately 25% lower bills and 10× faster codebase exploration for coding agents.

How It Works

SubQ's sparse-attention architecture identifies only the fraction of token relationships that matter and focuses compute exclusively on those, rather than evaluating every possible pair. Developers access SubQ via an OpenAI-compatible API endpoint, submitting full repositories or pipeline states in a single call with streaming and tool use support. Coding agent users install SubQ Code with a one-line command, and it automatically redirects expensive model turns to SubQ for token-heavy queries.

Use Cases

• Backend engineers processing full Python or JavaScript repositories through the API to get codebase-wide answers in one call.
• Coding agent developers integrating SubQ Code into Claude Code, Codex, or Cursor to cut inference costs on large context lookups.
• Enterprise ML teams running multi-step agentic pipelines that require persistent state across long task histories.
• DevOps teams analyzing months of pull request history — for example, six months of React PRs — within a single prompt.
• AI infrastructure researchers evaluating sub-quadratic architectures as a replacement for transformer-based long-context models.

Why Choose This Product

SubQ is best suited for teams and agents that regularly hit the context limits or cost ceilings of existing frontier models like GPT-5 or Claude Opus when working with large codebases or long task histories. The product is currently in private early access, so availability is limited to approved API and Code preview participants.