Platform hubs
Every framework, model provider and hardware platform that matters, described without marketing: what you install, what it costs, where the sharp edges are, and what moved recently.
PyTorch
The default deep learning framework for research and most production training, governed by the PyTorch Foundation, with eager execution, torch.compile and first-class CUDA, ROCm and Apple Silicon backends.
JAX and the Google stack
Google's composable numerical framework (jit, grad, vmap on XLA) with Flax, Optax and Orbax on top, the native path to TPUs and the second choice for new research code after PyTorch.
Hugging Face
The Hub where open models, datasets and demos live, plus the Transformers, Datasets, PEFT and TRL libraries that most people use to load, fine-tune and share them.
OpenAI
The GPT model family behind ChatGPT and the API whose request format most other providers copy, plus the gpt-oss open-weight models.
Anthropic
Maker of the Claude models and the Claude API, with a Messages format built around content blocks, adaptive thinking and prompt caching, sold directly and through AWS, Google Cloud and Microsoft Foundry.
Google DeepMind and Gemini
Gemini hosted models through AI Studio and Vertex AI, the open-weight Gemma family now under Apache-2.0, and the TPU hardware and JAX stack underneath them.
Meta AI and Llama
The Llama models that made open weights mainstream, the newer Muse line split between an API and an 816 B open release, and the PyTorch lineage that still shapes the tooling.
Mistral AI
The Paris lab whose Apache-2.0 releases set the bar for permissive open weights, now selling a 1 T parameter Large 4 through its Studio API alongside Small, Ministral, Codestral, Voxtral and OCR models.
NVIDIA
The GPUs almost all training runs on, the CUDA toolkit and libraries that lock them in, and a growing software layer of TensorRT-LLM, NIM microservices and open-weight Nemotron models.
Apple: MLX, Core ML and Apple Silicon
Running and training models on Apple Silicon's unified memory with MLX, converting PyTorch models for apps with Core ML and the new Core AI framework, and calling the on-device Foundation Models from Swift.
Microsoft: Azure AI, ONNX and Phi
Microsoft Foundry as the cloud catalogue for OpenAI, Anthropic, Meta, Mistral and open models, ONNX Runtime as the cross-platform inference engine, and the MIT-licensed Phi small models.
The open-source ecosystem (vLLM, llama.cpp, Ollama, LiteLLM, LangChain, LlamaIndex)
The permissively licensed serving engines, local runners, API gateways and orchestration libraries that turn open weights into running services without any vendor.