Autoregressive decoding is serial. Each token waits for the previous one. Speculative decoding breaks the chain: a cheap model drafts N tokens, the expensive model verifies all N in one forward pass. When the draft is right you paid one big forward for N generations.
A 70B LLM sampling one token takes ~30 ms on an H100. A 3B draft model takes ~3 ms. If we let the 3B draft 5 tokens ahead, then run the 70B once to verify all 5, the total is 5×3 + 30 = 45 ms for up to 5 accepted tokens — versus 5×30 = 150 ms for straight-line generation. That is the full speculative-decoding pitch: trade a small amount of extra GPU memory (draft model) for 2–4× lower decode latency.
The trick has to preserve the distribution. Speculative sampling, introduced by Leviathan et al. (2023) and by Chen et al. concurrently, guarantees that the output sequence is identically distributed to what the big model would have produced on its own. No quality tradeoff. Just faster.
Four families of draft-verifier pairs dominate 2026 inference:
Vanilla speculative (Leviathan 2023). Separate draft model (e.g., Llama 3 1B) + verifier (e.g., Llama 3 70B).
Medusa (Cai 2024). Multiple decoding heads on the verifier predict positions t+1..t+k in parallel. No separate draft model.
EAGLE family (Li 2024, 2025). Lightweight draft that reuses the verifier's hidden states; closer acceptance rate than vanilla; 3–4× typical.
Lookahead decoding (Fu 2024). Jacobi iteration; no draft model required at all. Self-speculation. Niche but dependency-free.
Every production inference stack in 2026 ships speculative decoding by default. vLLM, TensorRT-LLM, SGLang, and llama.cpp all support at least vanilla + EAGLE-2.
Full Access lesson
Keep going. The rest of this lesson is one step away.
Unlock all 503 lessons, video, code, quizzes and every future update. Pay in your own currency with Paystack.