Research and updates from the Reflex team.
How we replaced a 400M-parameter vision encoder with a 4.4M-parameter alternative — 51x token compression, 7.3x faster inference, trained for under $100.