gpt-oss-20b-speculator.eagle3

Model Overview

• Verifier: openai/gpt-oss-20b
• Speculative Decoding Algorithm: EAGLE-3
• Model Architecture: Eagle3Speculator
• Release Date: 04/02/2026
• Version: 3.0
• Model Developers: RedHat

This is a speculator model designed for use with openai/gpt-oss-20b, based on the EAGLE-3 speculative decoding algorithm. It was trained using the speculators library on hidden states distilled from openai/gpt-oss-120b, using the inference-optimization/gpt-oss-120b dataset (a combination of Magpie-Align/Magpie-Llama-3.1-Pro-300K-Filtered and the train_sft split of HuggingFaceH4/ultrachat_200k). This model should be used with the openai/gpt-oss-20b chat template, specifically through the /chat/completions endpoint.

Use with vLLM

vllm serve openai/gpt-oss-20b \
  -tp 1 \
  --speculative-config '{
    "model": "RedHatAI/gpt-oss-20b-speculator.eagle3",
    "num_speculative_tokens": 3,
    "method": "eagle3"
  }'

Evaluations

Use cases

Use Case	Dataset	Number of Samples
Coding	HumanEval	168
Math Reasoning	math_reasoning	80
QA	qa	80
Question Answering	question	80
RAG	rag	80
Summarization	summarization	80
Translation	translation	80

Acceptance lengths

Use Case	k=1	k=2	k=3	k=4	k=5
Coding	1.65	2.06	2.29	2.42	2.52
Math Reasoning	1.69	2.15	2.43	2.64	2.75
QA	1.50	1.74	1.88	1.91	1.93
Question Answering	1.58	1.90	2.07	2.19	2.25
RAG	1.53	1.82	1.95	2.04	2.08
Summarization	1.56	1.89	2.01	2.13	2.18
Translation	1.62	2.00	2.22	2.37	2.44

Details

Configuration

• temperature: default (greedy)
• repetitions: 1
• time per experiment: 5min
• hardware: 1xA100
• vLLM version: 0.17.1
• GuideLLM version: 0.5.5

Command

GUIDELLM__MAX_CONCURRENCY=128 \
GUIDELLM__PREFERRED_ROUTE="chat_completions" \
guidellm benchmark \
  --target "http://localhost:8000/v1" \
  --data "RedHatAI/speculator_benchmarks" \
  --data-args '{"data_files": "HumanEval.jsonl"}' \
  --rate-type throughput \
  --max-seconds 300 \
  --output-path "gpt-oss-20b-HumanEval.json"

GuideLLM interface changed, so for compatibility with the latest version (v0.6.0), please use the following command:

GUIDELLM__PREFERRED_ROUTE="chat_completions" \
guidellm benchmark \
  --target "http://localhost:8000/v1" \
  --data "RedHatAI/speculator_benchmarks" \
  --data-args '{"data_files": "HumanEval.jsonl"}' \
  --profile sweep \
  --max-seconds 1800 \
  --output-path "my_output.json" \
  --backend-args '{"extras": {"body": {"temperature":0.6, "top_p":0.95, "top_k":20}}}'