**Selective** **Restraint** Outruns **Profligate** **Proposals**

You may already run Qwen 3.8 27B on a home rig. It answers with strong quality but types with visible delay. On Oct 20, the eagle.ai blog introduced EAGLE-4 as swifter than DFlash 2 and DSpark, headlined by a table with 3.59x over baseline decoding. That figure is genuine, yet the setting matters a lot. It reflects four B300 accelerators serving a lone user with randomness disabled. In that narrow context, the claim holds. Outside it, calibration and concurrency reshape outcomes fast.

**Academic** **Pedigree** Underpinning **Adaptive** Proposals

Credit belongs with the originators of this research line. EAGLE-4 comes from Eagle AI, led by Hongyang Zhang of Waterloo, coauthor of versions one through three. Earlier this year the same group shipped EAGLE 3.1 with vLLM support, so continuity is real. Built-in prediction heads in DeepSeek V4, Qwen 3 and Gemma 4 borrow the same broad design, which means many users already touched a cousin of this approach. GitHub history under SafeAILab/EAGLE traces that lineage from early releases under Apache terms, while the fresh 12-page study plus GitHub code under eagle-hq/EAGLE-4 adds drafters for Qwen 27B and Nemotron Super. Nvidia, AMD, vLLM and SGLang appear as launch partners.

One concept explains the split from rivals. A large setup turns slow because each fresh token forces a full read of 27B weights from memory. Arithmetic is swift, waiting dominates. Imagine a crowded kitchen with a head cook and a brisk aide. The aide prepares upcoming dishes in advance, the cook reviews the whole row at once, keeps correct plates and remakes the first miss. One review releases several plates, yet quality matches solo work by the cook because every item passed inspection. MTP, DFlash 2, DSpark and EAGLE-4 all use this verify-after-draft loop with lossless output. Rivals differ only in drafting style. DFlash 2 and DSpark prepare a fixed-size block together, while EAGLE-4 drafts stepwise and reconsiders before each increment .

EAGLE-4 gains not from sharper guesses but from restraint . The EAGLE Harness measures accelerator speed, counts active users and weighs draft confidence before extending depth. Quiet load invites deep drafts with branching options, heavy load triggers a short draft and early halt. Skipped work never wastes verification budget. The drafter itself stays tiny, one layer versus five in rival drafters, which cuts compute per guess by around 3.8x . Next saving targets vocabulary scoring. Qwen vocabulary spans about 248k entries, and full scoring consumed most draft time. EAGLE-4 builds a shortlist of 128 likely entries first, then scores only that subset. The step shrinks by more than half while the shortlist still holds the correct entry in about 98% of cases.

**Calibration** Plus **Concurrency** Redefine **Leadership**

Headline figures depend on calibration . At zero randomness, where the setup always takes the top token, EAGLE-4 reaches 3.59x for a lone user, ahead of DFlash 2 at 2.80x, DSpark at 2.57x and built-in MTP at 2.22x. At randomness level one, common for chat interfaces, EAGLE-4 falls to 2.59x while DFlash 2 holds 2.39x, so the margin narrows from 28% to around 8%. Randomness hurts every drafter because matches grow rarer. Coding and math benefit most from zero randomness since deterministic answers are desired there anyway. Scale changes the order again. With 512 concurrent users, baseline serves about 21k tokens per sec in total, EAGLE-4 serves about 28k for 1.34x, while DFlash 2 drops to 0.78x, DSpark to 0.67x and MTP barely breaks even. Fixed blocks keep proposing work with no spare verification capacity, while adaptive depth keeps EAGLE-4 ahead. Always ask for randomness level and user count alongside any speedup claim.

The study lines up four options from familiar to novel. Built-in MTP needs no extra download and already runs in llama.cpp and LM Studio. It hits 2.2x for a lone user but slides toward 1.6x at randomness level one because it drafts a fixed chain. DSpark pairs parallel block drafting with a small sequential head for cleanup, measured at 2.57x. Since the DSpark team ships no Qwen drafter, the test used a community port from RedHat. The arXiv report 2607.05147 documents that parallel plus sequential design in full detail. DFlash 2 from Inco arrived in August with single-pass diffusion drafting and remains favored for this Qwen setup. Eagle measured 2.8x, while the Inco blog claims 2.7x to 3.4x on Qwen 27B, so the top of that range brushes EAGLE-4 territory. EAGLE-4 itself posts 3.59x, about 864 tokens per sec versus 240 without assistance. It also holds across stacks, with 3.49x on a lone MI355X accelerator and 2.74x versus 2.41x for MTP on Nemotron Super. The Unsloth guide offers a practical path for local MTP in llama.cpp and LM Studio, useful while heavier options mature.

Caution is warranted because independent reproduction is still missing. Every comparative figure so far originates from Eagle or launch partners. Rival runs used an engine the study leaves unnamed, so Inco has not yet shown what DFlash 2 achieves on the same rig with tuned settings. Public recipes also trail the peak. GitHub code under eagle-hq/EAGLE-4 supports patched SGLang and vLLM builds, with the support page listing around 3.3x to 3.4x on SGLang and near 3x on vLLM for a lone user. The vLLM docs describe speculative decoding configuration and draft acceptance in detail, which helps interpret that gap. The docs are unusually open about limits, a positive signal. Until a neutral lab repeats the matrix, treat leadership over DFlash 2 as a strong claim with an open question .

You cannot deploy it at home tonight. Shipments cover patched SGLang, vLLM and TokenSpeed only, validated on B300 and MI355X data-center accelerators. No integration exists for llama.cpp, LM Studio, MLX or Ollama, and no open pull request was live at recording time. Weight burden is steep. The tested Qwen 27B setup at full 16-bit precision spans about 55GB across four accelerators, plus a 6.2GB drafter holding a copy of the huge vocabulary. On a 16GB Mac mini, the drafter alone would consume more than a third of unified memory before the base weights load. Licensing adds friction. EAGLE-4 is source-available rather than fully open, free for personal work and research, free commercially only below $20M yearly revenue, with an extra clause around competing inference frameworks. That structure could delay a community port where most home users run versions today. Earlier EAGLE versions one through three used Apache terms, and EAGLE-3 reached llama.cpp in June, so the contrast is stark .

The most exciting datum hides outside the data center. A compact Gemma version in 4-bit ran on a Motorola Razr 70 Ultra using only handset CPUs via llama.cpp on a stock device. Baseline decoding managed about 23 tokens per sec on a coding task, EAGLE-4 reached about 83 for 3.6x . Energy fell by more than half, 166 millijoules per token versus 356, and the same answer completed in 14s rather than 43s. The Gemma drafter needs only 134MB, and a small four-core cloud node with 8GB of RAM averaged 2.7x, where EAGLE-3 managed 1.25x to 1.65x. Two caveats apply. The handset was cool at start and throttling begins after 30s to 45s of sustained output. More important, the tuned llama.cpp build and the tiny Gemma drafter remain unreleased, so the most home-friendly slice is not yet downloadable .

If you run Qwen 27B in llama.cpp or LM Studio, start with built-in MTP today. It costs nothing and already works. Next try DFlash 2 or DSpark, both bundled in LM Studio 0.4.22. Mac users can test a community MLX port that reached about 40 tokens per sec on an M4 Pro with 48GB, around 2.7x with DFlash 2 and DSpark close behind. Data-center teams with heavy concurrency and fitting license terms should trial EAGLE-4, since dense load is where adaptive halting shows the clearest edge. Wherever numbers appear, ask for randomness level and concurrent user count. A zero randomness lone-user score marks the ceiling, not the norm. In short, top speed comes from guesses wisely withheld , on a handset as reliably as inside a cluster.