BioCLIP 2.5

BioCLIP 2.5 is a CLIP-style vision foundation model for the tree of life: given a photograph of a plant, animal, or fungus, it produces embeddings that support zero-shot and few-shot classification of the organism's species and a range of other biological visual tasks. It is the successor checkpoint to BioCLIP 2 and was released in February 2026 by the NSF-funded Imageomics Institute, led by The Ohio State University. BioCLIP 2.5 keeps the training recipe and objective of BioCLIP 2 but scales the model up to a larger ViT-H/14 backbone and trains on an expanded version of the TreeOfLife-200M dataset.

The motivation is the same long-tailed problem that BioCLIP 2 targets: the space of life is vast, most species are rare and visually similar to close relatives, and labeled imagery is scarce for the majority of taxa. BioCLIP 2.5 addresses this by scaling both model capacity and training data. The image encoder moves from ViT-L/14 to the larger ViT-H/14, and the training corpus adds 19 million organism images to the original TreeOfLife-200M, alongside a roughly 26-million sample LAION-2B subset used for experience replay to retain general visual competence.

BioCLIP 2.5 is released as an open checkpoint on HuggingFace and shares its scientific description with the BioCLIP 2 paper (arXiv:2505.23883, a NeurIPS 2025 spotlight); it is a larger variant rather than a separately published model. As with BioCLIP 2, the central finding is that a narrow species-discrimination objective produces an embedding space with emergent structure useful well beyond species identification.

#Key Features

• ViT-H/14 backbone: BioCLIP 2.5 upgrades the image encoder from BioCLIP 2's ViT-L/14 to the larger ViT-H/14 Transformer, increasing model capacity while retaining the masked self-attention Transformer text encoder.
• Expanded tree-of-life training data: Training uses an updated TreeOfLife-200M corpus with 19 million additional organism images, paired with a LAION-2B subset for experience replay.
• Hierarchical contrastive learning: The model contrasts images against textual taxonomic hierarchies rather than flat class indices, encoding the nested structure of Linnaean classification, unchanged from the BioCLIP 2 recipe.
• Zero-shot and few-shot classification: The image-text embedding space classifies species never seen during fine-tuning simply by supplying candidate taxonomic or vernacular names, with no task-specific retraining.
• Consistent gains over BioCLIP 2: Scaling the backbone and data improves accuracy across biodiversity benchmarks without changing the objective.
• Open release: Weights are published on HuggingFace under the MIT license and load directly through the OpenCLIP library.

#Technical Details

BioCLIP 2.5 pairs a ViT-H/14 image encoder with a masked self-attention Transformer text encoder, following the BioCLIP 2 recipe: a hierarchical contrastive objective over full taxonomic labels, dual visual projectors, and an experience-replay mechanism that mixes in roughly 26 million LAION-2B samples. The model was trained on 32 H100 GPUs for 25 epochs over 11 days, using torch.compile and pure bf16 precision for acceleration. In the maintainers' evaluations, BioCLIP 2.5 raises zero-shot species-classification mean accuracy to 61.3% (a 5.7-point gain over BioCLIP 2's 55.6%) and mean performance on broader biological visual tasks to 61.0% (a 3.5-point gain over 57.5%), with an 8.7-point improvement on the FishNet habitat-distinction benchmark.

#Applications

BioCLIP 2.5 is aimed at ecologists, conservation biologists, systematists, and biodiversity-informatics teams who need to identify and characterize organisms at scale. Its zero-shot and few-shot capabilities suit camera-trap surveys, museum specimen digitization, citizen-science platforms, and rapid biodiversity assessment where labeled data for the target taxa is scarce. Because the embeddings capture ecological and morphological structure beyond species identity, they also support downstream tasks such as trait measurement, habitat inference, and organizing large unlabeled image collections, and the model can serve as a general visual encoder for other biological vision problems.

#Impact

BioCLIP 2.5 extends the BioCLIP line of biodiversity foundation models by demonstrating that scaling the backbone and training data continues to improve performance under the same hierarchical contrastive objective. Its MIT-licensed, OpenCLIP-compatible release lowers the barrier for the ecology and computer-vision communities to adopt a stronger organism-scale backbone as a drop-in successor to BioCLIP 2. As the larger member of the BioCLIP 2 family, it positions the Imageomics Institute's tree-of-life models as a widely reusable foundation for AI applied to the natural world.