Draft model
Small LLM that generates candidate tokens; larger "target" LLM verifies in parallel. Powers speculative decoding. Acceptance rate 50-80% → 2-3x speedup on inference without quality loss.
Classic pairing: Llama-3-8B as draft + Llama-3-70B as target. Draft generates 5-10 tokens forward; target verifies batch in one pass. If all match → accept all. If mismatch → accept up to divergence, regenerate from target. Self-speculative: target model itself is draft with truncated layers (Medusa, Eagle). Standard in vLLM, TGI, llama.cpp 2024+. Quality-neutral: target always determines final output, draft is just optimization.