Report
EAGLE-4 releases source-available code aimed at faster AI responses
The EAGLE-4 team reports 1.35× to 2.99× the throughput of matched no-speculation runs in its listed quickstarts.
TLDR
EAGLE-4's team released source-available software for speculative decoding, which uses a drafter to propose tokens for a larger language model to verify. The team says its strategy adapts to batch size, hardware and workload. Its listed quickstarts report 1.35× to 2.99× the throughput of matched no-speculation runs. The code is not OSI-approved open source, and some commercial uses require a separate license.
Combined views
31.4K
3 Sources, first seen ago
