Basically, the question is not whether you reach a particular percentage of sequence homology or sequence coverage: what you really need to know is whether, for classical homology modelling, you can define the correct sequence alignment between your query sequence and the modelling template. If you get this wrong, your model will be wrong. This is particularly difficult if there are a lot of insertions/deletions between the two sequences.

In respect to sequence coverage, this depends whether your protein consists of a single folding unit (domain) or of multiple domains: your template needs to cover the entire domain, but in principle you can model each domain of a multi domain protein from a different template. Some proteins also contain intrinsically flexible, natively unfolded segments that nevertheless can be important for function. There are programs that identify such segments based on their sequence properties (hydrophilic, high charge, high prevalence of Gly)

If you are working with a basic homology modelling program like Modeller ( https://salilab.org/modeller/download_installation.html ), you are yourself responsible for defining the correct sequence alignment. Beyond a simple sequence-sequence alignment, more advanced alignment programs take into account the evolutionary variability of each sequence : profile-to-sequence and profile-to-profile alignments that push the recognition limits to lower percentage of identical residues, and not only take into account whether two residues are identical, but also, whether amino acid substitutions are conservative (similar amino acids) or non-conservative. Amongst the more advanced homology modelling programs, SwissModel (https://swissmodel.expasy.org) does a pretty good job. Use the "search for templates" option first.Follow the links to the UniProt records of a list of potential templates - if these have a similar function to your query protein, this increases the probability that these are true orthologs.

In the twilight zone, where it becomes difficult to prove the correct alignment, you might try threading algorithms instead - these work not on a sequence- to- sequence alignment, but basically do the initial modelling and determine which of the models is most energetically feasible. The most widely used threading program is iTasser( https://zhanggroup.org/I-TASSER/ )

Where even this fails, where are the ab-initio fold prediction programs, e.g. Quark (https://zhanggroup.org/QUARK/), RosettaFold or DeepMind ( https://techcrunch.com/2021/07/15/researchers-match-deepminds-alphafold2-protein-folding-power-with-faster-freely-available-model/?guccounter=1&guce_referrer=aHR0cHM6Ly93d3cuZ29vZ2xlLmNvbS8&guce_referrer_sig=AQAAANRlv2RX-DB2rMCGK0k6Z6Ja_Har6_dG8irawCl7mnm0j9lQOO1rwhv_QH1VfWuxtYb0KBdIti9cQR5pLUoLJGiLheH-ujaebWoCfqo52GfYOBQPhfCgLjq0YrPa-npVNohBKKEEPIajL1AYdppmShYwngascd13HX4geWnCa_GX )