RedHatAI/Qwen3-235B-A22B-Instruct-2507-speculator.eagle3

Qwen30235B-A22B-Instruct-2507-speculator

Model Overview

• Verifier: Qwen/Qwen3-235B-A22B-Instruct-2507
• Speculative Decoding Algorithm: EAGLE-3
• Model Architecture: Eagle3Speculator
• Release Date: 12/03/2025
• Version: 1.0
• Model Developers: RedHat

This is a speculator model designed for use with Qwen/Qwen3-235B-A22B-Instruct-2507, 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. This model should be used with the Qwen/Qwen3-235B-A22B-Instruct-2507 chat template, specifically through the /chat/completions endpoint.

Use with vLLM

vllm serve Qwen/Qwen3-235B-A22B-Instruct-2507 \
  -tp 8 \
  --speculative-config '{
    "model": "RedHatAI/Qwen3-235B-A22B-Instruct-2507-speculator.eagle3",
    "num_speculative_tokens": 3,
    "method": "eagle3"
  }'

Evaluations

Use cases

Use Case	Dataset	Number of Samples
Coding	HumanEval	168
Math Reasoning	gsm8k	80
Text Summarization	CNN/Daily Mail	80

Acceptance lengths

Use Case	k=1	k=2	k=3	k=4	k=5
Coding	1.85	2.52	3.06	3.41	3.71
Math Reasoning	1.85	2.54	3.10	3.55	3.88
Text Summarization	1.61	1.98	2.15	2.24	2.29

Details

Configuration

• repetitions: 1
• time per experiment: 10min
• hardware: 8xA100
• vLLM version: 0.11.2
• GuideLLM version: 0.3.0

Command

GUIDELLM__PREFERRED_ROUTE="chat_completions" \
guidellm benchmark \
  --target "http://localhost:8000/v1" \
  --data "RedHatAI/speculator_benchmarks" \
  --data-args '{"data_files": "HumanEval.jsonl"}' \
  --rate-type sweep \
  --max-seconds 600 \
  --output-path "Qwen235B-HumanEval.json" \

</details>