jayyun98/Qwen3.8-27B-FP8-EAGLE4

Draftfit Qwen3.8-27B-FP8 draft — EAGLE4 compatible

Created with Draftfit, a platform for training, adapting, and validating speculative decoding drafts. This experimental checkpoint pairs the original EAGLE4 architecture with Qwen3.8-27B-FP8.

Measured acceptance

Draft	Accepted / proposed	Acceptance
Official EAGLE4 checkpoint	1,242 / 1,907	65.13%
This checkpoint	1,250 / 1,854	67.42%

Same 24 held-out prompts, 64-token greedy trajectories, seven-token proposals, and candidate-head rescoring against the same native FP8 target. These are cached-trajectory replay measurements. The gain is +2.29 percentage points; its paired prompt-bootstrap 95% interval is -0.96 to +5.55 points and includes zero. This small screening run does not establish a general improvement. Live speculative verification, exact output parity, and production serving speed remain unmeasured.

Use

hf download jayyun98/Qwen3.8-27B-FP8-EAGLE4

Use an EAGLE4-compatible runtime with this draft and the FP8 target above. The exported configuration and tensor names preserve the official EAGLE4 layout. A deployment command is not provided because native production serving has not yet been validated for this export. See evaluation.json and reproduction/ for the measured protocol and recipe.

Attribution and license

This checkpoint is derived from Eagle 4 technology developed by Eagle AI Corp. It is not an official Eagle AI Corp project. Original model: EAGLE-Corp/Qwen3.8-27B-EAGLE4.

This material is subject to the Eagle 4 Source-Available License, which is not an OSI-approved open-source license. Downloading, copying, accessing, modifying, using, hosting, distributing, or otherwise exercising licensed rights constitutes acceptance of that license and its incorporated trademark policy. See LICENSE, NOTICE, TRADEMARKS.md, and MODIFICATIONS.md.