Ohio State and NVIDIA Unveil BioCLIP 2 Biology Foundation Model at NeurIPS

Details

• BioCLIP 2 is a biology-focused foundation model trained on TreeOfLife-200M, a dataset with 214 million images across more than 925,000 taxonomic classes—the largest such dataset to date.
• Developed by Ohio State University's Imageomics Institute under Tanya Berger-Wolf, in collaboration with the Smithsonian Institution, the model was trained on 32 NVIDIA H100 GPUs over a 10-day period.
• It demonstrates emergent abilities beyond its original training goals, such as differentiating between species traits, analyzing inter- and intra-species relationships, and detecting organism health.
• Open-sourced on Hugging Face, BioCLIP 2 has reached 45,000 downloads in the past month, reflecting rapid adoption in research and an 18.1% accuracy improvement over its predecessor.
• The model will be presented at NeurIPS conferences in Mexico City and San Diego, with future plans to create wildlife digital twins for ecosystem simulation and public engagement platforms like zoo exhibits.

Impact

BioCLIP 2 marks a leap forward for domain-specific AI models, using vast, curated data to address key conservation and biodiversity challenges. By overcoming data scarcity that has long hampered species tracking, it provides a critical tool for both researchers and conservationists. Its public release and NeurIPS showcase are set to make this a pivotal resource for advancing ecosystem monitoring and environmental science innovation.