SubQ provides a large‑language‑model API and a plug‑in layer that can process up to 12 million tokens in a single prompt using a sub‑quadratic sparse‑attention architecture. This design reduces attention compute by up to 1,000×, delivering roughly 64× less compute and 56× faster performance than dense attention models while maintaining reasoning and knowledge capabilities, enabling developers to run full‑repo or long‑running agent contexts at linear cost.
Funding
Funding not disclosed
Founders
Product
Problem
Current large language models are constrained by quadratic attention, limiting context windows to a few thousand tokens and causing high compute costs for longer inputs. This restricts applications that need to process entire codebases, extensive documents, or long-running agent state in a single request.
Solution
SubQ provides a large language model built on a fully sub‑quadratic sparse‑attention architecture that can handle up to 12 million tokens in one prompt. By focusing computation only on the most relevant token relationships, the model reduces attention compute by nearly 1,000×, delivering roughly 64.5× lower compute than dense attention and 56× faster inference than FlashAttention‑2 at 1 M‑token contexts. The service offers an OpenAI‑compatible API with linear‑cost scaling, streaming support, and tool use, as well as a plug‑in layer for coding agents that redirects expensive model turns automatically. This enables developers to run full‑repository analyses, long‑document retrieval, and persistent‑state reasoning without sacrificing speed or cost.
Target Audience
Primary customers are developers and enterprise teams building AI‑powered applications that require processing of large codebases, extensive documents, or long‑running agent state, as well as providers of coding assistants and other long‑context agents.
Features
- Sub‑quadratic sparse‑attention architecture supporting a 12 M token context window
- Near‑linear cost scaling for full‑context API calls
- 64.5× lower compute than dense attention; 56× faster than FlashAttention‑2 at 1 M tokens
- OpenAI‑compatible endpoints with streaming and tool‑use capabilities
- One‑line install plug‑in for coding agents (Claude Code, Codex, Cursor) that auto‑redirects expensive turns
- High‑precision long‑context retrieval with near‑perfect single‑fact and multi‑task performance
- Efficient multi‑modal inference that maintains reasoning quality across large contexts