LangSplat constructs a 3D linguistic field by mapping the CLIP language into a set of 3D Gaussian distributions, enabling open-vocabulary queries of 3D scenes. It eliminates the costly rendering process found in NeRF, significantly enhancing efficiency. The learned language features accurately capture object boundaries, providing an accurate 3D linguistic field without the need for post-processing. LangSplat achieves a 199-fold speed improvement over LERF.