In speculative decoding a small, fast model guesses several words ahead and the large model checks them in one step. That makes answers faster.
Example
The small model suggests “thank you for your”, and the large model confirms it at once instead of word by word.
For pros
Because the large model checks the drafts in parallel and keeps only correct ones, the output stays equivalent. The gain depends on how often the small model is right. Many providers use this internally without showing it.
