Deployment & Inference · Level 4 of 5
Speculative Decoding
Accelerating generation by proposing tokens with a draft model and verifying them with a target model.
Correct algorithms can preserve the target sampling distribution while reducing costly target steps.
Example
A small model drafts several tokens that a larger model checks together.
Listen to the definition and example
Audio transcript
Speculative Decoding. Accelerating generation by proposing tokens with a draft model and verifying them with a target model. Correct algorithms can preserve the target sampling distribution while reducing costly target steps. For example: A small model drafts several tokens that a larger model checks together.
Explore this concept
Why it matters
This helps you balance prediction quality with memory, latency, and throughput.
Quick recall question
Try answering before looking back at the definition.