AnalysisDevelopersOctober 21, 2026

Tsinghua open-sources AKV, a KV cache server for LLM agents

Read original source →arxiv.org

AKV drops selected KV cache entries instead of summarizing history, cutting average cost 27.7% and lifting average task score 1.7 points across 3 models and 4 agent benchmarks. On SWE-bench Verified with GPT-OSS-120B, cost per task fell from 20.0 to 9.7 cents and score rose 40.0 to 43.2.

2 sources

More stories today

Open the live feed