§ Best of · Updated May 2026
Once an AI feature has users, infrastructure matters more than demos: latency, routing, observability, cost controls, model fallback, and deployment ownership. These picks cover the practical path from calling frontier APIs to serving open models at production scale.
§ The picks
The central hub for AI models, datasets, Spaces, libraries, and open-source ML collaboration.
The open-model hub: models, datasets, Spaces, inference endpoints, and the community graph that powers discovery.
Run open and community AI models from a web playground or API.
Fastest path from model demo to hosted API for image, video, audio, and open-source model experiments.
Fast generative media APIs for images, video, audio, and creative model workflows.
Generative media infrastructure when speed and API ergonomics matter for image and video products.
Serverless AI infrastructure for running code, jobs, containers, and GPUs from Python.
Python-native serverless GPUs for batch jobs, inference endpoints, and AI backend work without cloud ceremony.
Production AI inference platform for deploying, optimizing, and scaling models.
Production inference platform for teams serving custom and open models at scale.
One API and routing layer for hundreds of AI models across many providers.
Model router for comparing and switching providers without rewriting your app around every API.
§ Related recipe
Ship model-powered features without betting on one provider.
§ Common questions
Start hosted unless cost, privacy, latency, or control forces the issue. Self-hosting pays off later, but it adds operational work immediately.
Routers make it easier to compare models, fail over when a provider has issues, and optimize cost by sending easy tasks to cheaper models.
Cost per successful task. Token cost alone misses retries, latency, failures, and the human time spent fixing bad outputs.
§ More best-of lists