From a handful of conversations I had with people in HFT, researchers develop models in more conventional languages like java or python. The decision boundaries of these models are then translated into FPGA instructions for speed.