ABot-Recon released on HuggingFace: streaming 3D reconstruction from video

On HuggingFace, on August 26, 2026, the model acvlab/ABot-Recon was published. The authors describe it as a streaming 3D reconstruction model that estimates camera motion and scene geometry online from long videos, keeping only a local context of 12 frames in memory. The model predicts a point map in the current camera coordinate system and the relative pose of the neighboring frame, then aggregates local predictions into a global reconstruction. The weights are released in safetensors format, the largest file being 3.7 GB, with an apache-2.0 license. The code is open source; the engine will add the repository link.

What it means

In our reference on image-to-3d, we feature OpenLRM, Hunyuan3D 2.1, and TRELLIS. All three are 3D generators that create models from a single image. ABot-Recon solves a different task: it reconstructs scene geometry and camera trajectory from a video stream, rather than generating an asset from a picture. There is no direct overlap in use case with our tool pages. In terms of weight, ABot-Recon is comparable to Hunyuan3D 2.1, whose largest weight file is 6.9 GB, but it is noticeably heavier than TRELLIS (1.1 GB) and OpenLRM (1.7 GB). The apache-2.0 license matches the license of OpenLRM’s code and is more permissive than the cc-by-nc-4.0 for OpenLRM’s weights. VRAM requirements and the running platform are not specified in the model card — the developer does not clarify. For those looking for a tool to generate 3D models from a single image, ABot-Recon is not suitable: it is a model for scene reconstruction from video. If the task is to recover geometry from a long video recording, this model is worth considering, but hardware requirements will need to be checked in the repository.