Report
Qwen3-4B AI filter’s theoretical floor: about 6.6 seconds for 5,000 movie reviews on one H100
Full Stack Data Lab says per-row API calls miss query-planning benefits in batch work.
TLDR
Full Stack Data Lab estimates a theoretical minimum of about 6.6 seconds for one Qwen3-4B AI filter over 5,000 movie reviews on one H100. When sharing the lab’s blog in October 2026, the author said no system, including its open-source AI-SQL engine Quail, then came close. The author argues that per-row API calls miss query-planning benefits in batch work; the blog also examines how filter order changes the calculation.
Combined views
1 Source, first seen ago