The principle behind speculative decoding is simple. A small model drafts a continuation, then the large model verifies it in a batch. The more tokens it accepts, the more work you save.
The coordination problem
Sequential drafters build proposals one token at a time: each prediction informs the next, but every token adds a step. Parallel drafters propose several tokens together. They are fast, but individually plausible predictions can form an inconsistent sequence, and once verification rejects a token the rest is discarded.
What DSpine changes
DSpine, released by Tencent, moves that coordination into the layers of the network. At each layer, every position passes predicted token features to its neighbour, so all positions advance in parallel while information about preceding tokens flows through the depth of the network. The result is longer accepted sequences and more output from each expensive verification pass.
In Lisa
This matters for local inference, where reading the weights of a 27-billion-parameter model is expensive. We have integrated DSpine into Lisa, our Apple Silicon inference engine, in place of our sequential approach. Our tests show a 2.7 to 3.8 times speed-up over standard decoding, in line with the published range, and it is now the default drafter for Qwen3.8-27B models in Lisa.
