SubQ claims 128x compute savings at 2M tokens with Subquadratic Sparse Attention
ryanallred · x · 2026-10-13
Long-context startup SubQ details its SubQ 1.1 Small model (early access), built for "whole artifact reasoning" — ingesting an entire GitHub repo, years of SEC filings, all vendor contracts, or long-horizon agent state in one prompt, with no chunking or compression.
Approach: a proprietary Subquadratic Sparse Attention (SSA) algorithm that isolates the token relationships that matter, claiming 128x less compute than frontier models at 2M tokens. The company has shipped three model generations since May and is in private beta with some of the world's largest enterprises.
Context-scale references from the site:
- Five 500-page enterprise master agreements ≈ 1.6M tokens
- Three years of a Fortune 500 company's 10-K/10-Q filings ≈ 2.9M tokens
- The entire Python 3.13 standard library source ≈ 5.1M tokens
Target use cases: whole-repo code analysis, long-horizon agents that collapse brittle multi-agent orchestration into one coherent run, and trusted retrieval across a company's entire IP and documents.
More from Infra
- Seesa launches on-premise enterprise AI assistant deployed in US and UK firms — Scobleizer · 2026-10-13
- AI Token Economics Debate: Positive Inference Margins vs. the Shale-Oil Bear Case — toptickcrypto · 2026-10-13
- Perplexity built Projects' filesystem as a Git forge for agents, with greps beating semantic search — perplexity_ai · 2026-10-13
- Claude Code wiped a production database via Terraform, destroying 2.5 years of data — aakashgupta · 2026-10-13
- Six Five Pod: Microsoft-NVIDIA PC agents, Marvell's $70B-$90B FY2031 target — pstAsiatech · 2026-10-13
- 800 Blackwells vs 100,000: can labs without GPUs survive the open-weights race? — altryne · 2026-10-13