The purpose of LOD
Models in games consist of multiple versions of the same model with a different level of detail (LOD). High quality models with a higher polygon count and more details are used while a model is close to the viewpoint, lower quality models with a lower polygon count when it is further away. Having more polygons rendered in the viewpoint of the player results in a higher workload for rendering, causing a lower performance. For vehicles in GTA V, the highest quality model (L0) is included in the name_hi.yft file, lower quality models (L1 - L4) are included in the name.yft.
LODs in GTA V and their polygon count
The game's original car models from R* have, depending on their amount of details, roughly the following polygon counts:
L0: 50k - 60k
L1: 10k - 20k
L2: 6.000 - 11k
L3: 2.500 - 7.000
L4: 300 - 900
An example of an L0, L1 and L3 model.
If a model is more detailed than a default car, the L0 and the L1 are very likely to have more polygons than the default cars. The game is able to handle L0 models with more than 100k polygons, which is a good guidance level for more detailed vehicle models. For exceptionally detailed models, having an L0 with even more polygons is also usually acceptable without any noticeable hit on performance. The L1 model of a detailed, yet optimized vehicle is most likely going to have a polygon count of around 30k to 50k polygons.
Higher distance LODs, especially the L2 and above, should have a polygon count as close as possible to the default models. It's important to realise that the low quality models, especially the L3 and L4, are only used for very far distances. Therefore, there is no point in using high quality models. The details will not be visible, yet there most likely will be an impact on performance, especially when the car is used by NPCs in traffic.
The difficulty of creating LOD models is to create low poly models without sacrificing visual quality. The visual jump between LODs in-game should be as little as possible.
More information can be found in the guide Common Issues and Avoiding Them.
How to lower the polygon count of a model (L1 - L3)
A model's polygon count can be lowered by removing details. It's possible to remove parts completely (for example small screws or similar parts), but for most parts, it will be necessary to lower the polygon count of the original objects.
Welding in ZModeler 3
You select the group of vertices you want to connect and use the Modify > Submesh > Weld tool. This step is repeated until the details you wanted to remove is completely gone.
Using ZModeler's Optimize Mesh tool
You select a group of polygons (or even the whole model) and use the Modify > Submesh > Optimize polygonal mesh tool. You can experiment with the settings to see which result is the best.
This option is a lot faster than welding, however depending on the geometry, the result might not look the way you expected it. It's possible that the normals are wrong afterwards. In that case, it is recommend to lower the polygon count manually.
In some cases, an optimized model's normals might not be correct anymore. A guide explaining how to calculate normals can be found here: Calculating Normals
Creating the L4 model
The L4 model has such a low polygon count that welding is usually a worse option than to create a new model with low detail. The fastest way however is to find a similar looking default car and use its L4 model, if required with small edits and different textures.