Edit profile

AI settings

Available identifiers

Modular platform: each identifier is a plugin with its own capabilities, cost, and license. The cascade engine picks the right one automatically per observation.

Loading…

On-device AI (experimental)

Phi-3.5-vision (~4 GB) for offline ID + Llama-3.2-1B (~880 MB) for translations. WebGPU required. Models download once on first use, then run fully on-device.

Allow on-device AI as a fallback when no Claude key is set

Phi-3.5-vision is a general AI without taxonomic training. Results are capped at low confidence and never count toward research-grade.

Phi-3.5-vision

—

Llama-3.2-1B

—

BirdNET-Lite (audio, on-device)

Cornell Lab's bird-call identifier (~50 MB). Downloads once, then runs entirely on this device — recordings never leave the browser.

BirdNET-Lite v2.4

—

Cite Kahl et al. 2021 if you publish using BirdNET results.

BirdNET model is licensed CC BY-NC-SA 4.0 — non-commercial use only.

EfficientNet-Lite0 (photo, on-device)

Compact ImageNet classifier (~2.8 MB). Downloads once, then runs entirely on this device — photos never leave the browser.

EfficientNet-Lite0

—

Model: TensorFlow Model Garden (Apache-2.0). ImageNet labels are English common names — treat results as hints, not research-grade IDs.

Offline maps — Mexico

Download a Mexico-wide pmtiles archive (zoom 0–10) so the basemap renders without a network connection. Stored in your browser's Cache storage and reused on every map load.

Offline maps — Mexico

—

Edit profile

AI settings

Available identifiers

On-device AI (experimental)

BirdNET-Lite (audio, on-device)

EfficientNet-Lite0 (photo, on-device)

Offline maps — Mexico

Security