MASt3R = DUSt3R + a head that outputs a descriptor per pixel, trained so matching pixels have similar descriptors. Matches become pixel-accurate.
It also keeps the scale in the loss when the training data is metric ($z := \bar{z}$), so its pointmaps are approximately metric.
MASt3R-SLAM = L04's keyframe SLAM loop with MASt3R as the front end: track against the last keyframe, fuse, close loops. 15 frames per second, no calibration needed.

Two photographs of the same large country house taken from very different positions, joined by coloured lines between matched pixels found by MASt3R.

Figure: MASt3R repository, Naver Corporation, CC BY-NC-SA 4.0