Zyphra's ZHA matches dense models on SWE-bench and beats them on LongBench v2

We test ZHA on reasoning, coding, and long-context tasks: AIME 2026, SWE-bench Verified, Terminal-Bench 2.1, GPQA, RULER, and LongBench v2.

Deep ZHA matches dense on SWE-bench. On LongBench v2, it outperforms dense while attending to ~1.1% of decode keys.