When do SLMs suffice? Benchmarking task complexity, model scale, and architecture for water quality assessment

The rapid growth of water environment monitoring data has created pressing needs for efficient assessment and decision support. Large Language Models (LLMs) offer new opportunities for water quality assessment, yet how models of different parameter sizes and architectures compare in this domain has not been systematically studied. This study constructs a benchmark of 270 tasks across 9 complexity groups, proposes a subjective complexity metric integrating data, task, and output dimensions, and introduces the Normalized Time-Score (NTS) to jointly quantify report quality and processing efficiency on a common scale. Experiments using a Qwen3 controlled group (six models spanning from 1.7B to 235B) and a multi-source generalization group (six models spanning dense, MoE, and hybrid architectures) reveal that small Language Models (SLMs) achieve comparable quality to LLMs on simple tasks with substantially faster response, but LLM advantages expand progressively on complex tasks as SLM performance degrades through capacity-driven modular collapse. Beyond parameter scale, architectural design shapes this trade-off: selective-activation architectures exhibit greater performance stability than dense models of comparable or larger nominal size, shifting the SLM-to-LLM crossover point depending on relative architectural efficiency. Among the five evaluation dimensions, Clarity―rather than Completeness or Accuracy―emerges as the dominant performance differentiator under high-dimensional data conditions. Building on these patterns, a complexity-adaptive routing strategy that assigns tasks to SLMs or LLMs based on estimated complexity consistently outperforms both single-model baselines and cascading approaches. These findings provide an empirical basis for complexity-guided model selection and deployment in water environment monitoring scenarios.