PyTorch

Meta's PyTorch remained the dominant deep learning framework with 82% market share among AI researchers and engineers, used to train virtually every major AI model including GPT-4, Claude, Llama, and Stable Diffusion. PyTorch 2.2's torch.compile feature achieved 2x training speedups on H100 GPUs without code changes, and its adoption in production deployments via TorchServe grew 140% year-over-year.