KAITO (Kubernetes AI Toolchain Operator) is a Kubernetes operator from Microsoft that makes serving open-weight LLMs on a cluster a one-CRD operation. You write a Workspace resource that names a model — falcon-7b, llama-2-13b-chat, phi-3, mistral-7b, etc. — and KAITO takes care of provisioning GPU nodes, pulling a pre-built inference image, and exposing the model behind an OpenAI-compatible HTTP endpoint.
Under the hood it has two controllers. The workspace controller reconciles model workloads: it picks the right inference runtime (vLLM or Hugging Face TGI), mounts the model weights (either pre-baked into images hosted in MCR, or downloaded at runtime), and wires up a Service. The node provisioner integrates with Karpenter to request the exact GPU SKU the chosen model needs (A100, H100, specific VRAM class) from the underlying cloud, so you are not pre-sizing a node pool. KAITO also supports fine-tuning jobs using LoRA/QLoRA with the same Workspace abstraction, and RAG pipelines via a companion RAGEngine CRD.
It was donated to the CNCF sandbox in 2024 and is the opinionated “just serve this model” counterpart to lower-level stacks like KServe + vLLM. If you are on AKS the Karpenter integration is native; on other clouds you bring your own node provisioning.