gemma4-31B-it-speculator.eagle3

This is a preliminary model release, we will continue to train the model and improve the acceptance rates in the next few days.

Model Overview

Verifier: google/gemma-4-31b-it
Speculative Decoding Algorithm: EAGLE-3
Model Architecture: Eagle3Speculator
Release Date: 04/09/2026
Version: 1.0
Model Developers: RedHat

This is a speculator model designed for use with google/gemma-4-31b-it, based on the EAGLE-3 speculative decoding algorithm. It was trained using the speculators library on a combination of the Magpie-Align/Magpie-Llama-3.1-Pro-300K-Filtered dataset and the train_sft split of the HuggingFaceH4/ultrachat_200k dataset. Training data used Magpie + UltraChat with responses from the gemma-4-31B-it model (no reasoning). This model should be used with the google/gemma-4-31b-it chat template, specifically through the /chat/completions endpoint.

vLLM version

UPDATE: Now supported on vllm-main!

Use with vLLM

vllm serve google/gemma-4-31b-it \
  --tensor-parallel-size 2 \
  --speculative-config '{
    "model": "RedHatAI/gemma-4-31B-it-speculator.eagle3",
    "num_speculative_tokens": 3,
    "method": "eagle3"
  }' \
  --no-enable-prefix-caching \
  --max-num-seqs 64 \
  --enforce-eager

Evaluations

Model / run:
vLLM: UPDATE: Now supported on vllm-main!

Training data: Magpie + UltraChat; responses from the gemma 4 31B it model (no reasoning).

Use cases

Use Case	Dataset	Number of Samples
Coding	HumanEval	164
Math Reasoning	math_reasoning	80
Question Answering	qa	80
MT_bench (Question)	question	80
RAG	rag	80
Summarization	summarization	80
Translation	translation	80

Acceptance lengths (draft length, temperature=default)

Dataset	k=1	k=2	k=3	k=4	k=5
HumanEval	1.86	2.55	3.10	3.50	3.80
math_reasoning	1.87	2.59	3.15	3.59	3.93
qa	1.64	2.01	2.22	2.34	2.38
question	1.73	2.21	2.53	2.71	2.83
rag	1.72	2.21	2.50	2.65	2.80
summarization	1.60	1.92	2.07	2.15	2.20
translation	1.69	2.13	2.41	2.57	2.68