Zyphra ZHA Makes Long Context Subquadratic

// 1h agoRESEARCH PAPER

Zyphra ZHA Makes Long Context Subquadratic

Zyphra’s new research paper introduces ZHA, an attention mechanism that organizes recursively compressed KV-cache summaries into a tree and routes queries top-down to a sparse subset of tokens. The approach targets subquadratic scaling for long-context LLMs, with potential applications to million-token contexts.

// ANALYSIS

ZHA is compelling because it treats long-context attention as an indexing problem, but its biggest performance claims remain projections until tested at truly massive context lengths.

–Hierarchical routing avoids paying dense-attention costs across every token.
–A frozen backbone plus short KL-distillation training could make adoption easier than retraining new foundation models.
–The paper reports near-dense quality on retrieval, reasoning, math, and agentic coding evaluations.
–The most important engineering test is whether routing overhead, cache storage, and hardware behavior preserve the theoretical gains beyond 1M tokens.
–If validated, ZHA could make entire codebases, corpora, and agent histories practical inside one context window.

// TAGS

zyphra-hierarchical-attentionlong-contextllminferenceresearchtraining

DISCOVERED

1h ago

2026-11-02

PUBLISHED

1h ago

2026-11-02

RELEVANCE

9/ 10

AUTHOR

mark_k

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

MODEL30m ago

Deep Principle launches Hydrogen-1.0 science model

Deep Principle has released Hydrogen-1.0, a 753B open-weight model post-trained from GLM-5.3 for chemistry, materials science, and life-science research. It offers a 1M-token context window, MIT-licensed weights, and specialized training on laboratory records and discovery trajectories.

BENCHMARK37m ago

Prime Inference tops Nemotron cache rankings

Prime Intellect reports a 96.9% cache hit rate serving Nemotron 3.5 Super VL, cutting effective input costs to $0.024 per million tokens—roughly 70% below its largest provider rival.

INFRA52m ago

OpenAI's Pando Cuts Git Push Latency

OpenAI built Pando, a new Git backend for its monorepo that preserves existing workflows for engineers and coding agents. It reportedly cuts push latency by about 80% while serving nearly a petabyte of Git data daily.