The out-of-memory exception likely occurs because the vertex buffer must be backed by system memory in order to handle device loss in the underlying D3D9 implementation of XNA (it's probably in the
Transferring the data to the GPU and filling a vertex buffer with the data are the same thing (a vertex buffer is the only place in XNA's API you can store this data).
You're going to have to chunk the model file up into multiple vertex buffers when you load the data off the disk. The best way to do this is probably not to the naive thing and chunk it up sequentially, but rather to partition the model using some kind of spatial partitioning scheme (octree, bsp, what have you) so that you can completely avoid issuing draw calls for chunks of the model that are outside the view frustum.
This approach will:
Alleviate (but not solve) your system memory issues by requiring smaller allocations to hold the system-backed resources (even though the total memory usage will be the same or larger).
Improve your rendering performance by allowing you to make only the needed draw calls for only the needed parts of the model, based on frustum culling techniques. If you aren't clear on what frustum culling is, that might be a good topic to research or ask a new question about.