Serving a model is the part nobody puts on the keynote slide, right up until the bill arrives. On October 20 Google Cloud, Red Hat, and IBM put a 1.0 on llm-d, their shared project for running large models across a cluster.
What a 1.0 is asking you to trust
A 1.0 from three vendors is a support statement. The software is for the distributed problems: spreading a model that does not fit on one box, keeping the GPUs busy, failing over when a node dies. Red Hat’s customers already run Kubernetes this way. IBM’s customers already ask who answers the phone. Google Cloud wants the serving layer to be something an enterprise architect can approve.
It is not a model. Installing llm-d does not give you Gemini, and it does not answer a prompt. It gives the platform team a place to put whatever weights they are allowed to run.
The week it landed in
DeepSeek open-sourced DeepInfer a day earlier, aimed at its own V4.1 Flash model and at two specific chip families. llm-d is the more general offer: bring your orchestrator, bring a model you have rights to, serve it like any other serious service.
If you are choosing, pick by the cluster you already have. A new serving project does not pay for itself because the version number looks finished. It pays for itself when your current stack drops requests at the hour your users actually show up.