Match EAGLE-3 and Medusa-2 to drafting mechanism, integration cost and acceptance behavior

Classical speculative decoding uses a small draft model to propose tokens that the large target model then verifies. EAGLE-3 and Medusa-2 are 2026-era variants that skip the separate draft model entirely and instead attach drafting machinery to the target itself. **Medusa-2** adds several extra decoding heads to the target model. Each head predicts one of the next several positions in parallel: head 1 predicts t+1, head 2 predicts t+2, and so on. The target model then verifies all those guesses in one forward pass. Because the heads share the trunk, integration is straightforward and you do not have to ship and maintain a separate model. **EAGLE-3** drafts from the target's hidden states using a small auto-regressive head. Instead of guessing tokens directly, EAGLE looks at the rich internal features the target produces and uses those to predict the next few tokens with its own tiny autoregressive loop. Because the draft has access to feature-level information, its guesses match the target more often, so the target accepts more tokens per verification step. The practical tradeoff in 2026: EAGLE-3 has higher acceptance rates and bigger end to end speedups; Medusa-2 is easier to bolt on and tune. Both verify token trees (multiple candidate continuations) in a single target forward pass to maximize accepted tokens per pass.

EAGLE-3 and Medusa-2 are two members of the 2026 generation of self-drafting speculative decoding: instead of a separate small model proposing tokens, both attach drafting machinery to the target model. The differences come down to where the drafting happens and what the verification tree looks like. **Medusa-2** bolts several lightweight decoding heads onto the target. Head k is a small MLP that, given the last hidden state, predicts the token at position t+k. With four heads you propose four future tokens in parallel from a single target hidden state. The target then verifies them in one forward pass. Medusa-2 improved on Medusa-1 with better head training (joint training with the trunk) and tree-based verification (proposing a small tree of continuations rather than a single chain), which raises acceptance. **EAGLE-3** drafts at the feature level. A small autoregressive head reads the target's hidden states and runs its own short autoregressive loop, producing a tree of token proposals. Because the draft head sees the same internal features the target uses to choose tokens, its proposals are far closer to what the target would have produced, and acceptance rates climb. EAGLE-3 improves on EAGLE-2 with better tree construction and dynamic tree expansion (the draft tree is shaped by uncertainty rather than a fixed structure). Verification is symmetric across both: the target processes the proposed token tree in a single forward pass and walks down the tree accepting tokens as long as each one matches the target's sampling distribution. Higher acceptance means more tokens get verified per target pass, which is what produces the wall-clock speedup. EAGLE-3 reports higher acceptance and bigger speedups; Medusa-2 reports lower complexity, easier integration, and a more modest but still substantial speedup. Production stacks often expose both as plug-in choices.