Zyphra Hierarchical Attention (ZHA) is a tree-indexed sparse attention method that retrofits frozen LLMs with 4B tokens of distillation, matching dense attention on reasoning and 256K retrieval benchmarks.
The post Zyphra AI Introduces Zyphra Hierarchical Attention (ZHA): A Path to True Subquadratic Attention appeared first on MarkTechPost.