sampler in general (Drum Sampler too like BFD) & for “classic” instruments : Omnisphere/trillian. but I’m sure there is other synths if they use samples 
let say kontakt (sampler) let you choose between streaming or RAM like Omnisphere & trilian
It really depends on the VST and the programming.
Good modern samplers (ie higher end ones like Halion 4) usually offer disk streaming (now that we have faster drive controllers and drives) .
A VST that performs modeling of sounds will probably not use disk streaming as it’s just changing the waveform of the incoming audio signal.
A synth VST is (in most cases*) not going to use disk streaming because it is simply producing a waveform.
*(you can have ones, like Padshop, that do synthesis based on the waveform of a sample, but that’s usually one sample only).
I’ve seen virtual instruments that implement both. The Grand 3 will be my example.
TG3 allows for disk streaming.
It has a feature that reads the MIDI data and will unload (from RAM) unused samples.
not really still use ram but just for the begining of the sample !
You’d know the answer if you spent more time reading the other forums rather than posting nonsense here. 
it’s for streaming all the streaming goes for let say 10% ram before reading the disk
Everything has to go through RAM sometime.
Disk streaming will just use less RAM as it can replace the samples in RAM quicker.
yes it was just a explanation of disk streaming computer does’nt work without ram !
And I expanded on the foundation of your comment.
Didn’t mean it as a negative response to your comment.
Teamwork!!! 
Mostly true for modeling types, I would say.
The streaming sample players would not relieve “all” of your RAM, but can naturally use much much less RAM than a non-streaming version. It would also depend on the sample sizes used, how much of your resources would be saved. If a whole lot of smaller samples are used (but uses more than would be deemed loadable, so that streaming is elected instead) then you would probably waste valuable processing power instead, since it would be streaming samples that could have been loaded in their entirety without loosing too much RAM. It could be a fine line for some.
An example of modeling efficiency is AAS. They are known for their fast loading time and even faster switching of programs. Instead of using a lot of RAM, they are slightly heavier on processing since they build (in RAM of course, but uses very little) the sound on-the-fly, so to speak. Check out some of their plugins: http://www.applied-acoustics.com/