AlphaFold2 does not have 100 different trained models in the sense implied by num_models=100. The usual AlphaFold2 setup has five model parameter sets, so changing model_order to 1..100 will not create 100 valid model types.

If your goal is to sample more possible conformations, the usual approaches are:

1. Run the same 5 models with multiple random seeds.
2. Use ColabFold/local ColabFold options such as multiple seeds, if available in the notebook/version you are using.
3. Optionally enable stochastic/dropout-style sampling in workflows that support it.
4. Keep the outputs separated by seed/model and compare pLDDT, PAE, ranking confidence, and structural clustering.

For example, conceptually, 20 seeds x 5 AlphaFold model parameter sets gives 100 output structures. That is different from setting num_models=100.

A few cautions:

• The 100 structures are not 100 independent experimental observations.
• If the MSA/template signal is strong, many outputs may be nearly identical.
• If the protein has flexible regions or multiple domains, inspect the PAE plot; low pLDDT or high inter-domain PAE usually means uncertainty, not necessarily an alternative biological state.
• Do not pick a model just because it looks nicer in PyMOL; use the confidence metrics and, ideally, compare against known domains, homologs, mutations, or experimental constraints.

If you only need more models for visualization, multiple seeds are fine. If you need evidence for alternative conformations, AlphaFold alone is usually not enough; you would want orthogonal support such as experimental data, MD, known conformational states, or comparison to homologous structures.