This is a speculator model designed for use with Qwen3-30B-A3B with reasoning enabled. If interested in non-reasoning use cases, please use RedHatAI/Qwen3-30B-A3B-Instruct-2507-speculator.eagle3.

This model is based on the EAGLE-3 speculative decoding algorithm. It was trained using the speculators library on a combination of the Magpie-Align/Magpie-Pro-300K-Filtered and the HuggingFaceH4/ultrachat_200k datasets. The model was trained with thinking enabled. This model should be used with the Qwen3-30B-A3B chat template, specifically through the /chat/completions endpoint.

Use with vLLM

bash

vllm serve Qwen3-30B-A3B \
  -tp 1 \
  --speculative-config '{
    "model": "RedHatAI/Qwen3-30B-A3B-speculator.eagle3",
    "num_speculative_tokens": 5,
    "method": "eagle3"
  }'

Evaluations

Model / run: Qwen3-30B-A3B-speculator.eagle3 vLLM: 0.15.0
Training data: Magpie + UltraChat; responses from the Qwen/Qwen3-235B-A22B model (with reasoning enabled).

Acceptance lengths (draft length)

Dataset	k=1	k=2	k=3	k=4	k=5
HumanEval	1.81	2.44	2.90	3.21	3.44
math_reasoning	1.84	2.50	3.02	3.41	3.70
qa	1.69	2.15	2.44	2.61	2.72
question	1.76	2.32	2.71	2.93	3.09
rag	1.74	2.25	2.60	2.82	2.97
summarization	1.66	2.05	2.30	2.43	2.51
translation	1.72	2.21	2.53	2.74	2.87

Details

Configuration

Model: Qwen3-30B-A3B
Data: Magpie + UltraChat — responses from Qwen3-30B-A3B model (reasoning)
temperature: 0.0
vllm: 0.15.0
backend: vLLM chat_completions
rate-type: throughput
max-seconds per run: 300
hardware: 8× GPU (tensor parallel 8)
Benchmark data: RedHatAI/speculator_benchmarks
vLLM serve: --no-enable-prefix-caching, --max-num-seqs 64, --enforce-eager

Command

bash

GUIDELLM__PREFERRED_ROUTE="chat_completions" \
GUIDELLM__MAX_CONCURRENCY=128 \
guidellm benchmark \
  --target "http://localhost:8000/v1" \
  --data "RedHatAI/speculator_benchmarks" \
  --data-args '{"data_files": "HumanEval.jsonl"}' \
  --rate-type throughput \
  --max-seconds 300