We present a new jackknife estimator for instrumental variable inference with unknown heteroskedasticity. It weighs observations such that many-instruments consistency is guaranteed while the signal component in the data is maintained. We show that this results in a smaller signal component in...