Deploying open-weight models

How to actually run a downloadable model yourself, once *Closed-weight and open-weight models* has made the case to try: the serving stack a checkpoint sits inside, choosing a model and license operationally, sizing GPU memory against parameter count and quantization, what an inference runtime does, a hands-on local deployment, batching and the KV cache as the real capacity ceiling, scaling across accelerators, hardening a runtime into a production API, securing the model supply chain, the privacy duties that move onto you, evaluating and observing what you actually deployed, safe upgrades and rollbacks, and when managed inference beats doing it yourself.