Caffe
Berkeley released a library in which a network architecture is described in a text file and trained weights can be downloaded and used at once.
Why it matters
The first model zoo appeared: someone else's trained network could be taken and applied without repeating the training.
What mattered was not speed but the practice of sharing weights. Training AlexNet cost days on expensive hardware; Caffe let the result be published and taken. Fine-tuning on someone else's base grew out of that and became standard. TensorFlow and PyTorch displaced the library, but the habit of publishing weights remained.