Manycore Tech Open-Sources SpatialLM for Embodied Intelligence

Manycore Tech Inc. has announced the open-sourcing of its multimodal spatial comprehension model, SpatialLM, at the GTC 2025 event, as stated in a press release. This move is intended to lower barriers for training embodied intelligence by providing a foundational framework for practitioners to fine-tune the model for specific scenarios.

SpatialLM is designed to enhance robots' understanding of 3D environments by overcoming the limitations of traditional large language models. It can generate structured 3D scenes from videos using point cloud data, allowing for accurate recognition and translation into 3D layouts. Developers can access SpatialLM on platforms like Hugging Face, GitHub, and ModelScope.

Manycore Tech also highlighted its SpatialVerse platform, which complements SpatialLM by enabling the training of content-generation models in virtual settings. This combination provides a versatile system for digital simulations, enriching training data while maintaining high dataset quality. The company has established cooperation agreements with global players in the embodied intelligence sector to further this initiative.