KVarN is a native vLLM KV-cache quantization backend for your agents: 3-5x more context, throughput above FP16, and FP16-level accuracy. Calibration-free, one flag. Maintained by huawei-csl on GitHub, where it has earned 452 stars from the community.
