Spirit-v1.5

Spirit v1.5 is an open-source embodied-intelligence model developed by Spirit AI (千寻智能), a seed-stage portfolio company of Oasis Capital. Released in January 2026, it was the first Chinese open-source model to outperform Pi 0.5 on the RoboChallenge benchmark, ranking first overall in task score and success rate . The model uses a Vision-Language-Action (VLA) unified architecture and is distinguished by a training approach that deliberately avoids highly curated "clean" demonstration data, instead employing open-ended, diverse data collection oriented around "meaningful goal completion" to improve real-world generalization . Spirit AI co-founder Gao Yang described the underlying "data pyramid" philosophy: pre-training on massive human internet video rather than frame-by-frame world-model prediction, which the team argues cuts compute costs while achieving better results with fewer parameters . To validate its benchmark claims, Spirit AI open-sourced the base model weights, inference code, and usage examples alongside the results .