WildDet3D brings promptable 3D detection

This is the kind of release that pushes 3D perception from a narrow research demo toward a general-purpose spatial layer. The interesting part is not just open-vocabulary 3D boxes, but that Ai2 paired the model with enough data and deployment surfaces to make the category feel usable.

• –Text, point, and box prompts in one architecture make it easier to slot into VLM pipelines, trackers, and interactive AR apps.
• –The dataset scale is the real moat here: long-tail 3D detection usually breaks on coverage, and 1M+ images across 13.5K categories is a serious attempt to fix that.
• –Optional depth support matters because it makes the system more practical on devices that already have LiDAR, RGB-D, or stereo, instead of forcing RGB-only compromises.
• –The iPhone, Quest, and robotics demos signal a deployment-first story, not just a paper benchmark win.
• –For developers, the bet is clear: spatial intelligence is becoming a composable capability, not a single-task model problem.