Can Edge-Deployable Vision-Language Models Identify Species?

Ecological monitoring often relies on offline edge hardware, making small 2–8B vision-language models the primary candidates for field deployment. Evaluating models like Qwen3-VL and Gemma3 against the 300M domain specialist BioCLIP across camera-trap imagery reveals that general-purpose scale cannot substitute for specialized data: BioCLIP outperformed all tested VLMs by 33 to 59 percentage points. While all architectures degraded sharply under poor field legibility, open-set prompting also produced fabricated, taxonomically nonexistent species names in up to 9.6% of responses.