Speculative decoding: small draft model proposes tokens, big model verifies in one pass

HowDevelop · x · 2026-09-17

Local inference solves privacy and cost, but every token requires a full forward pass through the large model — a painful bottleneck for long-form or complex reasoning.

Speculative decoding changes the math: a smaller, faster draft model proposes 4-8 tokens ahead, the large target model verifies them all in one batched pass, accepting only tokens it would have generated anyway. Result: multiple tokens for roughly the cost of one forward pass, identical quality, faster output.

The catch: you can't pair any two models. Three things must align — a shared tokenizer (hard requirement), same model family (e.g. Qwen3-30B target + Qwen3-0.6B draft works), and a sensible size ratio.

Original post →

More from Infra

Infra channel →