Persistence of non-monotonic scaling beyond 8B parameters

Determine whether the observed non-monotonic relationship between Qwen3-VL parameter count and species-identification accuracy persists in Qwen3-VL variants with 32 billion or more parameters.

Background

Within the tested Qwen3-VL models, the apparent advantage of the 2B model over the 4B model in the original sample disappeared after expanding the sample, while the 8B model retained an advantage. The authors therefore characterize the observed scaling behavior as non-monotonic within the tested range but leave unresolved whether this pattern extends to substantially larger Qwen3-VL models.

References

We stress this non-monotonicity is established only within 2--8B; whether it persists at 32B or larger Qwen3-VL variants is an open question outside our scope.

— Can Edge-Deployable Vision-Language Models Identify Species?  (2609.11916 - Zhou et al., 10 Sep 2026) in Section 3, RQ2: family, prompting, and treatment, within the 2–8B range