jevIMAGE
23M-parameter Jev robotics model runs inference in 7ms
The author built a real-time phase-tracking model for robotics using the Jev approach, with 7ms inference that is 25x faster than SOTA and matches a 300x larger VLM on RoboProcessBench.
Belusochim Ugochukwu
@belusochimArchived 2026.10.03Originally posted 2026.10.03

Belusochim Ugochukwu @belusochimOriginally posted 2026.10.03
I built Jev for robotics. It’s a 23M parameter model for real-time phase tracking and subtask monitoring. - 7ms inference (25× faster than the current SOTA) - On par with a 300x bigger VLM on RoboProcessBench eval Designed to give policies fast, reliable progress signals without heavy compute overhead. More info here 👇🏽
x.com/belusochim/status/2106134977251549628Archived 2026.10.03 10:02