Researcher seeks feedback on Qwen2.5 memory limits for local deployment

A researcher is conducting a study to evaluate the memory limits of small AI language models (100M to 3B parameters) for retrieval and reconstruction tasks, aiming to identify those suitable for low-spec laptops. The study specifically benchmarks three Qwen2.5 models: 0.5B, 1.5B, and 3B.

• Retrieval memory performance was found to be relatively similar across all three model sizes.
• Reconstruction memory results showed an unexpected trend where the largest model, Qwen2.5-3B, performed the worst.
• The 3B model exhibited a significantly steeper decline in reconstruction performance as input information increased compared to the 1.5B and 0.5B models.

The author is requesting feedback on additional analyses or methodological checks to verify if this counterintuitive result is objective rather than an experimental error.