I’m experimenting with running DeepSeek locally and have seen speculative decoding suggested as a way to improve generation speed. Benchmarks are useful, but I care more about whether it makes interactive use feel snappier—especially on prompts where the model produces a longer answer.

For people who have tried it, did speculative decoding latency improve noticeably on your hardware? I’m also curious whether the gains depend heavily on the draft model or come with trade-offs that show up in answer quality, memory use, or setup time.