4GB is very tight for that vector set; you should have 8GB or more to load the full set. Alternatively you could use the optional limit argument to load_word2vec_format() to just load some of the vectors. For example, limit=500000 would load just the first 500,000 (instead of the full 3 million). As the file appears to put the more-frequently-appearing tokens first, that may be sufficient for many purposes.
No way to get away with 4G. I could load and compute GoogleNews-vectors-negative300.bin on my 8G RAM Macbook Pro. However, when I loaded this gigantic pretrained vector on AWS, I had to upgrade it to 16G RAM because it was serving a webapp at the same time. So basically if you want to use it on webapp with the safety margin, you would need 16G.
It is really difficult to load the entire Google-News-Vector pre-trained model. I was able to load about 50,000 (i.e 1/60 th of total) on my 8 GB Ubuntu machine using Jupyter Notebook. Yes, as expected, the memory/resource usage touched the roof.
So, it is safe to use atleast 16 GB to load the entire model, otherwise use limit=30000 as a parameter, as suggested by @gojomo.
