What you are experiencing is commonly referred to as “LOD,” or “level of detail.” This method lowers the visual quality of a model when a client is at a distance away from the model that doesn’t require as much detail. Some models, those of which are community made, don’t have LOD sub-models, and therefore can be very poorly optimized if many are on the map at the same time. TF2 models all have a minimum of three LOD models, the first being default, the other two being over some distances, and HL2 models of major characters have at least one second LOD model. This makes TF2 models really well optimized, but they aren’t so good for animations unless you take a few certain steps.
Ray Koeford, a pretty well rounded Garry’s Mod machinima maker from way back when, gives a few tips and tricks, one of which being how to remove LOD filtering, in a “tricks and tips” section of his site: