A comparative analysis of the performance between the large language models (LLM) Qwen3 and Qwen3.5, based on data aggregated from artificialanalysis.ai.

Comparison Methodology

The analysis distinguishes between dense models and Mixture-of-Experts (MoE) models. Dense models use their listed parameter size (e.g., 27B). For MoE models (e.g., 397B A17B), an effective size is calculated as the square root of the product between the total number of parameters and the number of active parameters. This conversion aims to provide an estimate of the compute-equivalent scale of MoE models, taking into account their specialized architecture.

For those evaluating on-premise deployments, there are significant trade-offs between dense and MoE models, particularly in terms of memory requirements and inference parallelization. AI-RADAR offers analytical frameworks on /llm-onpremise to evaluate these trade-offs.