Introduction

What Wingsmith is and is not, its pre-release status, and why the code still says hangar.

Wingsmith is a control plane for serving models across several Kubernetes GPU clusters. It enrols your clusters, keeps a per-device inventory of each, places each model on one or more clusters, renders it there as an NVIDIA Dynamo deployment, and routes OpenAI-compatible requests to it through its own gateway. Tenants, endpoints, quotas, and operations such as drain and scale are managed from a web console or the hgr CLI.

Wingsmith serves models. It does not run training or batch jobs. Inside each cluster it drives NVIDIA Dynamo, with vLLM, SGLang or TensorRT-LLM as the engine, and that is the only serving runtime it supports today.

Wingsmith is open source under the Apache 2.0 license and is developed by Aircloud. The project site is wingsmith.dev.

What it handles

• Cluster enrolment and per-device inventory
• Model placement: which clusters run which models
• Serving configuration and model lifecycle
• Quotas per project and device class
• An OpenAI-compatible gateway that routes each request to a cluster serving the model
• Operations such as drain, scale and cordon

The control plane stores intent (what you asked to run). Each cluster stores truth (what is actually running). See How it works.

What it is not

• A device scheduler. The scheduler inside each cluster picks GPUs. Wingsmith picks clusters.
• A Kubernetes dashboard. It works with your models, endpoints and quotas, not with CRD trees.
• A training or batch platform. It does not run jobs to completion and does not preempt.
• In the inference request path. Requests go through the gateway, which keeps serving when the control plane is down.

Status

Pre-release. There is no tagged release and no published images or chart. APIs, CRDs and the wire protocol change without notice.

Wingsmith is built for one organisation running its own clusters, operated by one admin or a small team sharing that role. There are no per-user roles and no SSO.

About the name

The product is called Wingsmith. The codebase is still called Hangar, and every code identifier keeps that name.

These docs say Wingsmith in prose and keep hangar and hgr in commands, configuration and API paths.

Where to go next