Nemo
Nemo (MAMBO_v3) identifies adult moths and butterflies in photographs, predicting 12,632 species, 4,476 genera and 104 families. Predictions use GBIF taxon IDs.
Try Nemo 路 Installation and API 路 Source code
Install with pip install 'mambo-v3[onnx,hub]==0.3.1':
from mambo_deploy import Predictor
model = Predictor.from_pretrained("asgersvenning/MAMBO-v3", backend="onnx")
prediction = model.predict("moth.jpg")[0]
print(prediction.label, prediction.confidence)
Performance
The comparisons cover three image domains:
- GBIF training domain: held-out test images from the same collection used to train Nemo (global scope; left panel).
- Automated monitoring: Danish AMI camera-light-trap image crops, expert-reviewed by Flemming Helsing (northern European scope; right panel).
- iNaturalist: recent research-grade images from its own platform, comparing Nemo, Nemo + TTA and Meghan with iNaturalist鈥檚 model (below).
Species-level scores after confidence filtering. Solid bars show full support; pale bars show shared support >5. Acceptance rates appear below each model.
Suggested species-confidence thresholds for Nemo ONNX (without / with TTA):
0.53 / 0.35 for general photographs (global scope);
0.82 / 0.75 for image crops of single individuals from camera light traps
(e.g. AMI; north_europe).
Recent research-grade iNaturalist images
941 adult-screened images with GBIF species labels from 1,000 recent Research Grade iNaturalist observations (2026-10-02). 890 reporting images; 51 separate calibration images. Global scope, no location input. Speed: Nemo ONNX and Meghan PyTorch on the same CPU. iNaturalist request time includes network latency; the pale bar shows the original 1,000-image collection average, including pacing, retries and other overhead.
Variants and details
- ONNX: CPU inference without PyTorch; the default Python backend.
- PyTorch: native CPU/CUDA inference through
mini_trainer, or the same portable API. - Scope: global by default; regional presets and custom species lists restrict predictions.
model="europe"selects geography, not the model generation. - Embeddings: 1,280 dimensions for downstream applications.
Nemo uses EfficientNetV2-S, trained for 30 epochs on 5,063,857 GBIF-sourced training images in the global-lepi collection. Training configuration 路 Provenance 路 Migration from Meghan (MAMBO_v2)
License
Model weights: CC BY-NC-SA 4.0 (attribution, non-commercial, share-alike). Code: MIT. See notices and attribution.