Taylor says Gemini beat stacked Jev on image detection

Nathan LeClaire posted a demo of Jev stacked on Jev, where the model divides an image up and conquers it section by section to find bounding boxes. He credits Lily for the inspiration. His own framing is not a victory lap. For straightforward detection he says it is worse and slower than SAM3, and the case he makes for it is the odd semantic query, his example being finding all the sheep who might "yee haw".

Mike Taylor ran the same kind of task through Gemini and said it just does a better job, and that because Gemini only has to make one call it ends up cheaper. He posted a Claude artifact explainer alongside the reply.

Mike Taylor

@hammer_mt

X

Because gemini only has to make one call it actually ends up cheaper

What the posts do not contain

No accuracy number. No price. No image count, no latency, no description of how either output was scored. This is one person looking at one set of results next to another and saying which he preferred, posted as a reply, not as an eval. Taylor did run it, which puts him ahead of the people with opinions and no output, but a reply with two screenshots is not a benchmark and nobody involved claimed it was.

The cost argument is the part that generalises without needing a scoreboard. Divide and conquer over an image means N calls where a single pass means one, and N multiplies whatever the per call price is. Jev's per call price has been the thing carrying it in image work so far, including the MotherDuck image comparison where the pitch was Sonnet quality for a fraction of the spend. Stacking throws that advantage away by design.

For people building on it, the interesting question is not which model won Taylor's image. It is whether your detection task actually needs the decomposition. Bounding boxes arrived as a first class type in Jev 1.14, so a single call is available, and if a single call clears your bar then the stacked version is paying for structure you are not using.

Nathan LeClaire

@dotpem

X

For straightforward detection it's worse/slower than SAM3