VLM that says "it's not there" is now open source.

From the source

Ask most grounding models for an object that isn't in the image, and you still get a box. WeVisOne-VL returns None. On our WeRef bench, the 4B rejects 96.4% of those queries. On Ref-ZOM, a public benchmark built for the same failure mode, it rejects 86.0%.

Now a box means something. One set of weights for five jobs that used to mean five models: detection, counting, OCR, masks, and alpha mattes from a 9.6M-param decoder. Alpha mattes are higher precision than binary masks: soft, per-pixel opacity, so hair, fur, and semi-transparent edges come through intact.

No SAM in the pipeline. Images only. Our scores, rerunnable with the repo's eval scripts: MixCount counting: 40.4 TotalText OCR: 52.0 P3M matting error, lower is better: 7.3 Dense200 pointing: 87.0 Full comparison tables in the README. Apache 2.0, 2B/4B weights and demo: https://t.co/qIRinDHXkl