# Zyphra's 6-level router reaches 93.3% on RULER-13 at 256K

> Original title: Crucially, deep trees actually work.

- Source: Zyphra
- Published: 2026-11-02T22:04:55.000Z
- Bad Influence: https://badinfluence.news/items/pie44q3ows0v1w80ghrge8rck
- Original: https://x.com/ZyphraAI/status/2117377439466934698

## Summary

Zyphra reports that deeper tree routers can retrieve better than shallow ones at the same compression across 28 tested geometries. Its 6-level router scores 93.3% on RULER-13 at 256K context, versus 96.0% for dense attention, while attending to just 2.8% of decode keys.

## Post

Crucially, deep trees actually work.

We tested 28 geometries. At fixed compression, deeper trees often retrieve better at the same sparsity.

Our 6-level router scores 93.3% on RULER-13 at 256K (vs 96.0% dense), while attending to just 2.8% of decode keys.

![](https://badinfluence.news/api/img-proxy?u=https%3A%2F%2Fpbs.twimg.com%2Fmedia%2FHWJsH9JaIAAXg9x.jpg&mode=thumb&exp=1793923200&sig=65b2218369ef52da)
