gemma-4-31b-it-heretic-ara-eagle3-ko

This is Eagle-3 draft model for Korean conversation only. For other languages, please use other model.

You may expect 1.5x speed boost at maximum on Korean workload.

한국어 대화에 최적화 된 Eagle-3 Draft 모델입니다. 타 언어의 경우 다른 모델을 사용하세요.

한국어 작업에서 최대 1.5x 정도 속도 개선이 있습니다.

Model Overview

• Verifier: hell0ks/gemma-4-31b-it-heretic-ara-FP8
• Speculative Decoding Algorithm: EAGLE-3
• Model Architecture: Eagle3Speculator
• Release Date: 2026/05/05

How it was made

• Training framework: Speculators
• Datasets: Private (Korean 8: English 2, 60k, No Reasoning)
• Training hardware: 1 DGX Spark

Name	Value
Learning Rate	1e-4
Scheduler Type	Cosine
Warmup steps	50
Sequence length	4096
Epochs	4
Vocab size	32000

Usage

Tested with vLLM on DGX Spark (sm121)

vllm serve hell0ks/gemma-4-31b-it-heretic-ara-FP8 --port 8000 --reasoning-parser gemma4 --enable-auto-tool-choice --tool-call-parser gemma4 --speculative-config '{"model": "hell0ks/gemma-4-31b-it-heretic-ara-eagle3-ko", "num_speculative_tokens": 3, "method": "eagle3"}'