Speculative decoding is using a small model to quickly generate a sequence that every so often you pass through a larger model to check and correct. It can be much faster than just using the larger model, with tolerably close accuracy.