SpatialLM is an open-source 3D large language model that processes point cloud data and produces structured scene understanding for indoor spaces. Developed by Manycore Research, the project was accepted at NeurIPS 2025 and has earned more than 5,000 stars on GitHub, with releases continuing into 2026.
Where ordinary vision-language models describe images in free text, SpatialLM outputs structured 3D models of rooms: architectural elements such as walls, doors and windows, together with oriented object boxes and categories. It takes raw point clouds, for example from RGB-D scans or LiDAR, and turns them into layouts that downstream software can use for reconstruction, robotics or interior design. The team publishes model checkpoints built on Llama and Qwen backbones through Hugging Face, including SpatialLM 1.1 versions with a stronger point cloud encoder and higher resolution, along with the SpatialLM dataset and fine-tuning instructions.
The project site collects demos, the technical report and links to all released checkpoints, and the repository documents data preparation, inference and evaluation in detail, so research groups can reproduce the published results and benchmark the model against their own indoor scanning data before building on it.
SpatialLM is a free research release. The code is available on GitHub and the trained models and dataset can be downloaded from Hugging Face at no cost, subject to the license terms stated in the repository.
SpatialLM is for researchers and engineers working on 3D scene understanding, indoor mapping, robotics and spatial intelligence. Developers building scanning, real-estate or design applications can use it as a foundation for converting point clouds into structured, machine-readable room models instead of training a perception stack from scratch.