Camera traps often operate on edge hardware in the field with limited or no connectivity, making small, locally‑deployable vision‑language models (VLMs) the practically relevant class to evaluate rather than frontier‑scale ones. We selected four VLMs in the 2–8 B parameter range (Qwen3‑VL 2B, 4B, 8B, Gemma3 4B) and compared them against the domain‑specific specialist BioCLIP (300 M parameters). The evaluation used a 96‑species task, testing clean iNaturalist photographs and camera‑trap imagery from six LILA.science collections, forming two independently sampled test sets. All models performed far above chance, but every model—general‑purpose or specialist—suffered a sharp drop on field imagery, with domain gaps of 9.6–26.6 percentage points that were consistent across taxonomic levels and both test sets. This indicates the degradation stems from reduced image legibility rather than a failure in fine‑grained discrimination. Despite its much smaller size, BioCLIP outperformed every VLM by 33.2–59.2 points on an expanded 200‑image sample per model, suggesting that specialized training data matter more than model scale. Yet BioCLIP’s own domain gap (18.0 points) is statistically indistinguishable from the best VLM’s (22.3 points), reinforcing that the clean‑to‑field performance loss is a property of the image‑quality shift. Under open‑set prompting, 5.9–9.6% of responses are syntactically valid but correspond to nonexistent species names, and the relative fabrication‑rate ranking across models replicates exactly across both evaluation sets, a more robust finding than any single accuracy estimate.
Review: The work highlights that in resource‑constrained field deployments, domain‑specific data outweigh sheer model size. Advancing edge VLMs for species identification will likely depend on gathering more high‑quality, domain‑focused training images rather than merely scaling up parameters.