Same topic, related formats. Practice these next.
EAGLE-3 drafts from the target model's hidden states with a small autoregressive head; Medusa-2 drafts with multiple parallel token-prediction heads sharing the trunk.
Two ways to guess what a careful writer is about to say next. Medusa-2 attaches a small panel of helpers to the writer's elbow, each calling out a different upcoming word. The writer then checks the whole guessed phrase at once. It is easy to bolt these helpers on, and they speed things up, but they sometimes guess the wrong words. EAGLE-3 listens directly to the writer's thoughts (the hidden state) and uses a tiny apprentice writer to draft the next few words in sequence. Because the apprentice has access to richer information, more of its drafts match the master writer, so the master accepts longer chunks at once. EAGLE-3 is harder to wire up but accepts more, so each verification step buys more tokens.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Everything important, quickly.
3 min: define self-drafting + Medusa-2 parallel heads versus EAGLE-3 autoregressive head over hidden states + why feature-level drafting raises acceptance + tree-based verification + the integration vs acceptance tradeoff.
| Property | Medusa-2 | EAGLE-3 |
|---|---|---|
| Drafting mechanism | Parallel lightweight heads on the trunk | Small autoregressive head over hidden states |
| Draft input | Last target hidden state (single point) | Sequence of target hidden states (feature-level) |
| Drafted output shape | Tree of token guesses (shallow) | Tree of token guesses (deeper, dynamic) |
| Acceptance rate | Moderate, falls off fast with depth | Higher, holds up at greater depth |
| Integration cost | Low: train heads, add scheduler | Higher: hidden-state plumbing plus AR draft head |
| Reported speedup in 2026 | 1.5-2.5x on chat decode | 2-4x on chat decode |
Real products, models, and research that use this idea.
What an interviewer would ask next. Try answering before peeking at the approach.
Red flags and common mistakes that signal junior thinking. Click to expand.
Calling Medusa and EAGLE 'small draft models'. Neither is a separate draft model in the classical sense; both attach drafting machinery to the target model itself. Medusa adds extra heads; EAGLE adds a small autoregressive head that reads the target's hidden states.
The night-before-the-interview bullets. Scan these on the way to the call.
Primary sources. Skim if you want the original framing.