Run Your Own Inference Stack on Vultr with Modelplane
Vultr Kubernetes Engine (VKE) is now a supported inference cluster provider in Modelplane, the open source control plane for AI models. As of the v0.3 release, you can describe a GPU cluster on Vultr in a short YAML manifest. You specify the region, node pool, instance class, and the minimum and maximum node counts you want it to scale between, and Modelplane has the full serving stack stood up for you: VPC, control plane, system and GPU node pools, and the inference layer on top.
Deploying a model afterwards is another manifest, and it is the same one you would use on any other provider.
Why Modelplane
Modelplane is an open source project that treats models, engines, clouds, and accelerators as one system you operate yourself, rather than a set of provider-specific consoles and scripts. Its provider model is deliberately uniform: one resource describes the hardware, another provisions it, and the same node pool shape applies everywhere.
Vultr now sits alongside GKE, EKS and AKS in that lineup. The practical upshot for Vultr customers is portability in both directions. The ML half of a deployment manifest doesn't change when the hardware underneath it does, so putting a workload on Vultr Cloud GPUs is a configuration decision rather than a migration.
Point Claude Code at your own Vultr Cloud GPUs
The other v0.3 change worth calling out for anyone running GPUs here: Modelplane now serves the Anthropic Messages API end to end. A vLLM server exposes the Messages endpoint alongside its OpenAI-compatible routes, and Modelplane's routing preserves the path below the service prefix, so a single service address answers both.
Any client that speaks the Messages API works against it, including Claude Code. Point it at your service address, map its Opus, Sonnet, and Haiku tiers onto the model name you're actually serving, and you have a coding agent backed entirely by a model hosted on your own Vultr cluster. Modelplane's Anthropic Messages API recipe covers the tool-calling flags that make agentic use work and why you'll want to cap output tokens on a smaller model.
Get started
Stand up an inference cluster on Vultr Kubernetes Engine with Modelplane's getting-started guide, or read the full v0.3 release notes. If your workload needs a provider or instance class that isn't covered yet, the project takes contributions and issues on GitHub.
