---
format: "aidr-story-markdown/v1"
id: "56a3a8ea11d2c429f9fb75ca41ec1f7c4f0b2fcc3991062d76e9ffeaa9a06795"
canonical_url: "https://aidr.today/56a3a8ea?lang=en"
title: "Zyphra Debuts ZHA Method to Reach Subquadratic AI Attention Speeds"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-11-02T22:38:04.000Z"
category: "Research"
topics: ["inference","llm"]
source_urls: ["https://marketbrief.now/ai/zyphra-debuts-zha-method-to-reach-subquadratic-ai-attention-speeds-eac3006c","https://huggingnews.com/ai/zyphra-debuts-zha-method-to-reach-subquadratic-ai-attention-speeds-eac3006c","https://www.marktechpost.com/2026/11/02/zyphra-hierarchical-attention-zha-sparse-attention/"]
summary: "The AI lab Zyphra has released a new computation method for treating long sequences of tokens known as Zyphra Hierarchical Attention (ZHA). The system utilizes a learned tree of compressed summaries to enable subquadratic attention, a technique that reduces the computational overhead required for models to process massive context windows. Zyphra projects that the ZHA method will be 35x faster than dense attention and 5x faster than Multi-head Sparse Attention (MSA) when processing 10M tokens. The company expects the performance gap to widen at 1B tokens, where it projects speedups of roughly 400x over dense attention and 55x over MSA."
---

# Zyphra Debuts ZHA Method to Reach Subquadratic AI Attention Speeds

> [Open the canonical story](<https://aidr.today/56a3a8ea?lang=en>)

**Published:** 2026-11-02T22:38:04.000Z
**Category:** Research
**Topics:** inference, llm

## Summary

The AI lab Zyphra has released a new computation method for treating long sequences of tokens known as Zyphra Hierarchical Attention \(ZHA\)\. The system utilizes a learned tree of compressed summaries to enable subquadratic attention, a technique that reduces the computational overhead required for models to process massive context windows\. Zyphra projects that the ZHA method will be 35x faster than dense attention and 5x faster than Multi\-head Sparse Attention \(MSA\) when processing 10M tokens\. The company expects the performance gap to widen at 1B tokens, where it projects speedups of roughly 400x over dense attention and 55x over MSA\.

## Sources

- [Story source](<https://marketbrief.now/ai/zyphra-debuts-zha-method-to-reach-subquadratic-ai-attention-speeds-eac3006c>)
- [Story source](<https://huggingnews.com/ai/zyphra-debuts-zha-method-to-reach-subquadratic-ai-attention-speeds-eac3006c>)
- [Story source](<https://www.marktechpost.com/2026/11/02/zyphra-hierarchical-attention-zha-sparse-attention/>)