AI Model
Summary
RedHatAI released a preliminary EAGLE-3 speculator model for Gemma-4-31B-it to accelerate inference via vLLM. Trained on Magpie and UltraChat data, it improves token acceptance rates across coding, math, and QA benchmarks.
Highlights
- Uses EAGLE-3 speculative decoding for Gemma-4-31B-it.
- Released by RedHat as a preliminary v1.0 model.
- Optimized for vLLM with num_speculative_tokens=3.
- Shows improved acceptance lengths in HumanEval and math reasoning.
auto-generated
Context
Audience
Machine Learning Engineers
DomainMachine Learning
Formatmodel repository
Accessopen source
Topics
Related
Google Gemma-4-31B-itEAGLE-3 AlgorithmvLLMSpeculative DecodingMagpie DatasetUltraChat Dataset