← Retour au blog
tech 17 June 2026

Diving into SubQ 1.1 Small: Revolutionizing Enterprise AI with Efficient Attention

SubQ 1.1 Small offers a breakthrough in processing long contexts with its subquadratic sparse attention model. With processing capability up to 12M tokens, it promises to transform how enterprises handle complex data.

Article inspired by the original source
SubQ 1.1 Small ↗ subq.ai

Introduction

In a world where businesses are flooded with complex data, finding efficient solutions to process this information is crucial. SubQ 1.1 Small emerges as a significant advancement in this domain, thanks to its subquadratic sparse attention model that promises to revolutionize enterprise data processing.

Context: The Challenge of Enterprise AI

The challenges of enterprise AI are often related to the ability to process complete artifacts such as entire codebases, document collections, or financial filings. Traditional solutions involve retrieval pipelines, chunking strategies, and agentic scaffolding, which, while useful, are merely workarounds for the contextual limitations of current model architectures.

SubQ 1.1 Small: A Technological Breakthrough

SubQ 1.1 Small is the latest Subquadratic Sparse Attention (SSA) model that overcomes the primary hurdle: the quadratic constraint of attention. It allows up to a 1,000x reduction in attention computation, maintaining execution speed 56 times faster than FlashAttention-2.

Key Features

  • Long-context retrieval up to 12M tokens: SubQ 1.1 Small excels in the "needle-in-a-haystack" test, proving its efficiency in locating facts buried deep within large contexts.
  • Optimization of context length and general reasoning ability: With strong performance across knowledge, coding, and non-coding enterprise agent benchmarks.
  • Computational efficiency: At 1M tokens, it reduces the need for computation by 64.5 times compared to dense attention.

Performance and Benchmarks

SubQ 1.1 Small was evaluated across five axes: long-context retrieval, context-length generalization, knowledge, coding, and long-horizon agentic tasks.

  • Long-context retrieval and generalization: With near-perfect scores at 1M, 2M, 6M, and 12M tokens, SubQ 1.1 Small demonstrates its ability to generalize attention based on content relevance.
  • RULER capability test: With a performance of 99.12% at 128K, it surpasses the demands of multi-faceted tasks such as variable tracing and frequency aggregation.

Conclusion

With SubQ 1.1 Small, Subquadratic offers a revolutionary solution for businesses seeking to optimize the processing of complex data. By removing the quadratic constraint, this model opens up new possibilities for enterprise AI.

Let's discuss your project in 15 minutes.

SubQ 1.1 Small Subquadratic Sparse Attention Enterprise AI Long-context retrieval Efficient computation
Deepthix newsletter · 100% AI · every Monday 8am

An AI agent reads tech for you.

Our AI agent scans ~200 sources per week and ships the best articles to your inbox Monday 8am. Free. One click to unsubscribe.

Visit the newsletter page →

Want to automate your operations?

Let's talk about your project in 15 minutes.

Book a call